Research Methodology and Data Analysis

No part of this digital document may be reproduced, stored in a retrieval system or transmitted in any form or
by any means. The publisher has taken reasonable care in the preparation of this digital document, but makes no
expressed or implied warranty of any kind and assumes no responsibility for any errors or omissions. No
liability is assumed for incidental or consequential damages in connection with or arising out of information
contained herein. This digital document is sold with the clear understanding that the publisher is not engaged in
rendering legal, medical or any other professional services.
Research Methodology and Data Analysis

E-Records Integrity Requirements

Orlando López
2022. ISBN: 978-1-68507-553-8 (Hardcover)
2022. ISBN: 978-1-68507-726-6 (eBook)

Predictive Analytics for Data Driven Decision Making – Tools

and Techniques for Solving Real World Problems
B. Uma Maheswari, PhD, L. Ashok Kumar, PhD
and R. Sujatha, PhD
2022. ISBN: 978-1-68507-650-4 (Hardcover)
2022. ISBN: 978-1-68507-770-9 (eBook)

Top 10 Challenges of Big Data Analytics

Maria José Sousa, PhD
and Manuel Au-Yong Oliveira (Editors)
2021. ISBN: 978-1-53619-133-2 (Hardcover)
2021. ISBN: 978-1-53619-189-9 (eBook)

Validation of Instruments for the Investigation of Sexuality

in Vulnerable Groups
Lubia del Carmen Castillo Arcos (Editor)
2020. ISBN: 978-1-53618-340-5 (Hardcover)
2020. ISBN: 978-1-53618-471-6 (eBook)

Statistical Modelling of Complex Correlated and Clustered

Data Household Surveys in Africa
Ngianga-Bakwin Kandala, PhD
and Lawrence Kazembe, PhD (Editors)
2019. ISBN: 978-1-53615-981-3 (Hardcover)
2019. ISBN: 978-1-53615-982-0 (eBook)

More information about this series can be found at

https://novapublishers.com/product-category/series/research-methodology-and-
data-analysis/
Cem Ufuk Baytar, PhD
Editor

The Future of Data Mining

Copyright © 2022 by Nova Science Publishers, Inc.
DOI: https://doi.org/10.52305/KCIN5931

All rights reserved. No part of this book may be reproduced, stored in a retrieval system or transmitted
in any form or by any means: electronic, electrostatic, magnetic, tape, mechanical photocopying,
recording or otherwise without the written permission of the Publisher.

We have partnered with Copyright Clearance Center to make it easy for you to obtain permissions to
reuse content from this publication. Simply navigate to this publication’s page on Nova’s website and
locate the “Get Permission” button below the title description. This button is linked directly to the
title’s permission page on copyright.com. Alternatively, you can visit copyright.com and search by
title, ISBN, or ISSN.

For further questions about using the service on copyright.com, please contact:
Copyright Clearance Center
Phone: +1-(978) 750-8400 Fax: +1-(978) 750-4470 E-mail: [email protected].

NOTICE TO THE READER

The Publisher has taken reasonable care in the preparation of this book, but makes no expressed or
implied warranty of any kind and assumes no responsibility for any errors or omissions. No liability is
assumed for incidental or consequential damages in connection with or arising out of information
contained in this book. The Publisher shall not be liable for any special, consequential, or exemplary
damages resulting, in whole or in part, from the readers’ use of, or reliance upon, this material. Any
parts of this book based on government reports are so indicated and copyright is claimed for those parts
to the extent applicable to compilations of such works.

Independent verification should be sought for any data, advice or recommendations contained in this
book. In addition, no responsibility is assumed by the Publisher for any injury and/or damage to persons
or property arising from any methods, products, instructions, ideas or otherwise contained in this
publication.

This publication is designed to provide accurate and authoritative information with regard to the subject
matter covered herein. It is sold with the clear understanding that the Publisher is not engaged in
rendering legal or any other professional services. If legal or any other expert assistance is required,
the services of a competent person should be sought. FROM A DECLARATION OF PARTICIPANTS
JOINTLY ADOPTED BY A COMMITTEE OF THE AMERICAN BAR ASSOCIATION AND A
COMMITTEE OF PUBLISHERS.

Additional color graphics may be available in the e-book version of this book.

Library of Congress Cataloging-in-Publication Data

ISBN: H%RRN

Published by Nova Science Publishers, Inc. † New York

Contents

Preface .......................................................................................... vii

Acknowledgments ....................................................................................... ix
Chapter 1 Data Analytics Applied to the Human
Resources Industry ............................................................1
Cem Ufuk Baytar
Chapter 2 Toxicogenomics Data Mining as a Promising
Prioritization Tool in Toxicity Testing ..........................13
Katarina Živančević, Dragica Bozic,
Katarina Baralić and Danijela Đukić-Ćosić
Chapter 3 Applications of Data Mining Algorithms
for Customer Recommendations
in Retail Marketing .........................................................29
Elif Delice, Lütviye Özge Polatlı,
İrem Düzdar Argun and Hakan Tozan
Chapter 4 Analysis of Customer Churn in the Banking
Industry Using Data Mining ...........................................51
Özge Doğuç
Chapter 5 The Crowdsourcing Concept-Based
Data Mining Approach Applied
in Prosumer Microgrids ..................................................77
B. C. Neagu, M. Gavrilaș,
O. Ivanov and G. Grigoraș
Chapter 6 Active Learning ...............................................................95
Jože M. Rožanec,, Blaž Fortuna
and Dunja Mladenić
vi Contents

Chapter 7 Prediction of General Anxiety Disorder

Using Machine Learning Techniques ..........................119
Kevser Şahinbaş
Editor’s Contact Information ..................................................................139
Index .........................................................................................141
Preface

The purpose of this book is to discuss data mining, which is a subset of data
science, from a variety of perspectives. With the technological advances of
recent years, new software and hardware-based systems are available in most
business environments. With these systems, data production continues to
increase in personal, corporate, commercial and many other areas. Information
systems convert raw data, which alone are not so meaningful, into information
after the processes are applied. Database systems are necessary for the storage
and management of the information generated. Revealing meaningful
relationships hidden in a stack of high-volume data shows the function of data
mining. Processing big data has become important to produce information that
will support business decisions and be a strategic tool in today’s competitive
environment. In this context, the effectiveness of data mining applications is
increasing day by day as a decision support system to develop marketing
strategies in every sector by identifying customer behavior and target groups.
The wide perspective of book chapters is likely to provide a rich source
of information for those interested in broadening their understanding of the
multiple facets associated with data mining. The primary audience for this
book includes business professionals, practitioners, data scientists as well as
researchers who dedicate their work to issues associated with data mining (or
data science). An important motivation for editing this book was the need to
create an organized framework for providing various points of view about the
future of data mining.
Chapter 1 provides details about the relationship between the concept of
analytics and the field of human resources. It also explains the opportunities
waiting for human resources. Chapter 2 presents toxicogenomic data mining
as a promising functional annotation-based prioritization tool, with a concise
summary of potentially useful software and tools, including their major
advantages and limitations. Chapter 3 introduces applications of data mining
algorithms (decision tree and random forest) for customer recommendations
in retail marketing. Chapter 4 is an analysis of customer churn in the banking
viii Cem Ufuk Baytar

industry by using different machine algorithms and creating a model on the

Knime platform. Chapter 5 discusses a data mining approach applied in
prosumer microgrids based on the crowdsourcing concept. Chapter 6 takes an
in-depth look at the field of active learning, which is a sub-field of machine
learning. Chapter 7 is about prediction of general anxiety disorder by using
machine learning techniques.

Cem Ufuk Baytar

Istanbul, Turkey

August 2022
Acknowledgments

I would like to convey my appreciation to all contributors. I also offer my

thanks to my colleagues and family whose support was a source of
encouragement during the period when I spent hours on this project.
It is my pleasure to thank all the professionals at Nova Science Publishers
with whom I enjoyed working.
Chapter 1

Data Analytics Applied to the Human

Resources Industry

Cem Ufuk Baytar

Management Information Systems, Istanbul Topkapı University,
Istanbul, Turkey

Abstract

The department of human resources in companies plays an active role in

the architecture of a strategic workforce in terms of creating and
managing human capital. It is necessary to embody its strategies with
appropriate raw data and accurate analyses. Human resources
management gains the feature of being a strategic partner in company
management by generating strategies based on consistent measurements
related to the workforce. As a result of such an approach, it has an
important place in creating competition in businesses and maintaining the
existence of businesses. Innovations, which emerge with the
transformation of data into information, not only offer a competitive
advantage to companies but also help them make better decisions. In
parallel with the developments in technology, changes in the business
environment force human resources management to be prepared for these
changes and to transform itself with this change. In this study, the concept
of analytics is discussed concerning human resources. In addition, by
trying to explain the relationship between human resources and analytics
concepts, the opportunities waiting for human resources are revealed. As
a result, it is clear that digital transformation and developments in
technology make the importance of human resources increase.

*
Corresponding Author’s Email: [email protected].

In: The Future of Data Mining

Editor: Cem Ufuk Baytar
ISBN: 979-8-88697-250-4
© 2022 Nova Science Publishers, Inc.
2 Cem Ufuk Baytar

Keywords: data analytics, data mining, data management, human resources

metrics

Introduction

With the developments in Information and Communication Technologies,

data can be stored and managed by bringing it together from different sources.
Today, it is possible for managers to reach and use big data with various types
of technologies at any time. Since deciding is one of the most critical processes
for companies, it is necessary to organize and analyze large amounts of
complex data by appropriate methods for decision-makers. In this context,
data analytics is a field that emerged in recent years that aims to convert raw
data into information with scientific and mathematical processes and methods.
It has become a very interesting topic both in academia and in the business
world in a short time. There are different definitions of human resources
analytics, i.e., it means using advanced statistical analyses, predictive
modeling procedures, and human capital investment analysis to forecast and
extrapolate “what-if” scenarios for decision-making (Falletta and Combs,
2021). This study has the following objectives: (a) improving the
understanding of data analytics, (b) identifying the role of data analytics in
human resources management, (c) assessing the practices of human resources
analytics, and (d) discussing the study’s implications and future research
directions

Data Analytics

The use of analytics can be traced back to the 1910s when scientific
management emerged (Mortenson, Doherty and Robinson, 2015). It can be
considered as the continuation of optimization and simulation techniques that
could achieve maximum output with limited resources in the 1940s (Delen and
Zolbanin, 2018). At the end of the 1960s, the use of computer systems to
analyze data and support decision-makers led to the use of analytics.
Applications called decision support systems were used in activities such as
product planning, investment portfolio management, and transportation route
(Davenport and Harris, 2007). These decision support systems were named
Operations Research and Management Science in general. In addition to the
Data Analytics Applied to the Human Resources Industry 3

advanced Operations Research and Management Science models that were

used in many industries and management systems in the 1970s, rule-based
Expert Systems (Expert Systems) emerged (Delen and Ram, 2018). Through
expert systems, direct use of computers and data involved monitoring and
reporting performance rather than decision-making (Davenport and Harris,
2007). In the 1980s, the application of multiple and separate information
systems gave way to integrated information systems at the enterprise level,
called Enterprise Resource Planning (ERP). Relational Database Management
took the place of the schemas that show the data and do not have a certain
standard (Delen and Ram, 2018). Thus, data from each unit of the enterprise
could be collected and brought together and each unit of the enterprise had
access to this data. In the 1990s, expert information systems emerged.
Business Intelligence and Information Technologies started to be used in
business environments (Krishnamoorthi and Mathew, 2018). Data could then
be stored centrally and holistically in data warehouses. In the late 2000s,
business analytics started to be used as a very important and analytical
component of Business Intelligence (Krishnamoorthi and Mathew, 2018).
In these years, Data/Text Mining, Cloud Programming, and SaaS
(Software as a Service) methods have also been developed (Delen and
Zolbanin, 2018). In the 2010s, the internet was largely used by mobile devices,
etc. Due to its widespread use with devices (Internet of Things), the new data
generation big data has started to develop rapidly. Although there are many
data sources in this period, social network/media data is seen as the most
interesting (Delen and Ram, 2018).
Data analytics includes a group of steps to find out trends and to get results
by processing data sets. To manage the process of data analytics, various types
of software and specialized systems have been used by specialists. In other
words, it is a field of analyzing and managing data to extract valuable insights
from data. Statistical analysis and technologies are applied in the field of data
analytics. It plays a vital role to analyze business processes and improving
decision-making to get a competitive advantage (Stedman, 2022). Data
analytics has relationships with many disciplines, for instance, statistics,
mathematics, and computer programming. To make data analysis, specialists
use different techniques. Some of them are data mining, data cleansing, data
transformation, data modeling, etc.
Data analytics has two general methodologies. Exploratory data analysis
tries to find patterns in data whereas confirmatory data analysis uses statistical
techniques to determine whether hypotheses related to the data set are true or
not. There are four types of data analytics as follows (Olavsrud, 2022):
4 Cem Ufuk Baytar

• Descriptive analytics analyses current or historical data by finding

patterns to describe the present situation or situation in a specified
time interval, i.e, view of business intelligence in business analytics.
• Diagnostic analytics focuses on data, generally produced by
descriptive analytics, to find out cause-effect relationships related to
past events.
• Predictive analytics aims to make predictions about future outcomes
by using some techniques, i.e., machine learning, deep learning,
forecasting, etc., based on data from descriptive and diagnostic
analytics. It is generally considered a type of advanced analytics.
• Prescriptive analytics is a type of advanced analytics. It makes a
suggestion of a solution to reach a specified target, for instance,
machine learning, business algorithms are used in business predictive
analytics.

Companies in every sector use data analytics to improve operational

efficiency and increase revenues. For example, UPS has developed the
Harmonized Enterprise Analytics Tool (HEAT). This tool helped to keep track
of the real state of every package in the system by analyzing customer data,
operational data, and planning data. In another example, Kaiser Permanente
aimed to improve daily operations efficiency and customer satisfaction. To
reduce waiting times in queues, the company used machine learning and
artificial intelligence to manage data operations in its 39 hospitals in 2015
(Olavsrud, 2022).
There are different data analytics methods and techniques as defined
below (Olavsrud, 2022):

• Regression analysis is used to understand statistically whether there

is a relationship between variables and to calculate the level of
relationship between them, i.e., how much social media activities may
affect the amount of sales.
• Monte Carlo simulation is often used for risk analysis. Specialists
make models calculate the probability of events that can not be
estimated because of the side effects of some variables.
• Factor analysis is a statistical method for massive data to break into
smaller parts to manage them easily. In this way, it enables to discover
the hidden patterns in a huge data set. For example, factor analysis is
beneficial to find out some facts about customer loyalty in business.
Data Analytics Applied to the Human Resources Industry 5

• Cohort analysis is a technique that is used to group datasets for

analyzing data groups having common characteristics (cohorts). This
is often used to understand customer segments.
• Cluster analysis is used to classify structures in a dataset into related
groups (clusters). For example, insurance companies try to
understand the reasons why some specified locations are related to
some types of insurance.
• Time series analysis is a statistical technique that tries to find out
trend analysis. Data in a series of particular time intervals is analyzed
to discover trends over time. It is useful for economic and sales
forecasting.
• Sentiment analysis aims to identify feelings transferred from a
dataset. Some tools are used, i.e., such as natural language processing,
text analysis, and computational linguistics. sentiment analysis tries
to interpret qualitative data, i.e., understanding how customers feel
about a brand.

The process of producing meaningful information is called information

discovery because of pre-processing large volumes of data after it is stored in
information systems. Data mining is one of the important steps referenced in
information discovery (Yıldırım and Birant, 2018). The steps, which are
included in the information discovery process, are listed as follows (Han and
Kamber, 2006):

• Data cleaning: Removal of data with the highest level of noise and
inconsistency
• Data integration: Connecting multiple data sources
• Data selection: Removing data, which need analysis, from the
database
• Data conversion: Data collection or conversion to appropriate formats
• Data mining: Creating a data pattern by using appropriate methods
• Pattern evaluation: Identification of patterns representing information
based on certain criteria
• Information representation: Representation of information by using
visualization techniques

Data mining consists of methods (models) that allow hidden and

meaningful patterns to be found in the data stack. These methods are divided
6 Cem Ufuk Baytar

into two groups. The first of these groups consists of predictive methods. In
the other group, there are descriptive methods. The main predictive methods
are classification and regression models. Descriptive methods perform
definitions of patterns hidden among the data. For example, revealing whether
there is a similarity between the purchasing habits of married men and single
women during shopping may be an example of the descriptive model.
Clustering methods and association rules are among the descriptive methods.
According to Özdemir et al. (2018), in predictive methods, a model is created
using data with all the details previously known. These model results are
applied to unknown data to achieve the results (Özdemir et al., 2018).
The classification method is the process of adding new data to
predetermined classes by applying the results to this model after the creation
of a model through known data (Silahtaroğlu, 2013). In this context, regression
analysis commonly used in classical statistics is a kind of classification
method (Gamgan and Altunkaynak, 2017). Only numeric values are used in
the regression model. Some of the classification and regression models are
decision trees, artificial neural networks, genetic algorithms, k-nearest
neighbor, and bayes classifiers (Cömert and Kaymaz, 2019).
Decision trees form a tree-shaped structure for decision-making (Sayad,
2022). Decision trees are made up of branches and leaves. In this context,
decision nodes can be divided into two or more branches. Decision nodes are
used for decision-making or classification purposes depending on the data set.
Leaf nodes in decision trees represent decisions. The root node is at the top of
the decision tree. To reach a decision, a specific path is traced from the root of
the tree to the leaf nodes. Algorithms such as ID3, C4.5, and CART are used
to create decision trees (Balaban and Kartal, 2018).
Clustering is the job of assigning data to groups of objects with similar
characteristics (Ceylan et al., 2017). This analysis provides the formation of
data groups with common attributes. This method plays a role in determining
the correct varieties. At the same time, clustering analysis is used for different
purposes such as hypothetical testing, prediction for groups of objects, and
detection of contrary values. The clustering of data contributes to the analysis
of the data, while it also causes the details to be overlooked. Each group has a
homogeneous structure in its own right (Berkhin, 2006). There are clustering
models such as K-means, Fuzzy C-Means (FCM), Kohonen Networks,
Hierarchical Clustering, and so on. The similarity between the data is
measured by measuring the distances of the data from each other.
One of the best-known methods for clustering is the k-means algorithm,
developed by Mac Queen in 1967 (Arai and Barakbah, 2007). This algorithm
Data Analytics Applied to the Human Resources Industry 7

is based on the idea that a midpoint can represent a cluster (Steinbach et al.,
2000). The algorithm is one of those clustering algorithms that automatically
create clusters from similar things and is one of those types of unsupervised
learning without any reference or preliminary information. It is called k-means
because it creates the k original set and the center of each set is the average of
the values in the cluster (Harrington, 2012).

Human Resources Analytics

Human resources (HR) analytics is human resource management associated

with data for creating views deeply on how investing in assets of human capital
affects success in business based on: (a) creating revenue, (b) minimizing
expenses, (c) decreasing risks, and (d) executing strategic plans. It is realized
by applying statistical methods to integrated human resources, talent
management, and financial and operational data. HR analytics specifically
deals with the metrics of the HR function, such as time to hire, training
expense per employee, and time until promotion. All these metrics are
managed exclusively by human resources (Lalwani, 2022). In other words,
human resource analytics is defined as the systematic identification and
measurement of human stimulations according to business results (Heuvel and
Bondarouk, 2016). HR analytics is segmenting the workforce and using
statistical analyses and predictive modeling procedures to identify key drivers
(i.e., factors and variables) and cause and effect relationships that enable and
inhibit important business outcomes (Falletta and Combs, 2021). The field of
converting human behavior into data and examining it is defined as people
analytics. Human resources analytics covers all business processes by starting
with human resources. It can also be defined as a communication tool that
combines data from different sources to describe the current situation and
predict future outcomes. (Fitz-Enz, 2010).
Several HR metrics contribute to business value. Based on the key
performance indicators (KPI) of the organization, HR can then propose the
metrics that can influence these KPIs. Common metrics tracked by HR
analytics are as follows (Lalwani, 2022):

• Revenue per employee is calculated by dividing a company’s revenue

by the total number of employees in the company. This metric
indicates the average revenue each employee generates. It is a
8 Cem Ufuk Baytar

measure of how efficient an organization is at creating revenue by

employees (Jayanthi, 2020).
• The offer acceptance rate is defined as dividing the total number of
job offers by the number of accepted formal job offers given in a
certain time interval. A higher rate (above 85%) means a good ratio.
If it is lower, this metric can be used to re-evaluate the strategy of the
company’s talent acquisition (Keçecioğlu and Oktay, 2010).
• Training expenses per employee are obtained by dividing the total
training expense by the total number of employees who received
training. The value of this expense can be determined by measuring
the training efficiency. Poor efficiency may lead HR management to
re-evaluate the training expense per employee.
• Training efficiency can be measured by the analysis of various data
sources, such as performance improvement, test scores, and upward
transition in employees’ roles in the organization after training.
Measuring training efficiency can be crucial to measure the
effectiveness of a training program.
• The voluntary turnover rate occurs when employees voluntarily
choose to leave their jobs. It is calculated by dividing the number of
employees who left voluntarily by the total number of employees in
the organization. This metric can lead to the identification of gaps in
the employee experience that are leading to voluntary attrition
(Dursun, 2013).
• Involuntary turnover rate means termination of an employee from the
related position. The rate is calculated by dividing the number of
employees who left involuntarily by the total number of employees
in the organization. This metric can be tied back to the recruitment
strategy and used to develop a plan to improve the quality of hires to
avoid involuntary turnover.
• Time to fill is the number of days between advertising a job opening
and hiring someone to fill that position. By measuring the time to fill,
recruiters can alter their recruitment strategy to identify areas where
the most time is being spent (Anger et al., 2021).
• Time to hire is the number of days between approaching a candidate
and the candidate’s acceptance of the job offer. Just like time to fill,
data-driven analysis of time to hire can benefit recruiters and help
them improve the candidate experience to reduce this time (Jayanthi,
2020).
Data Analytics Applied to the Human Resources Industry 9

• Absenteeism is a productivity metric. It is measured by dividing the

number of days missed by the total number of scheduled workdays.
Absenteeism can offer insights into overall employee health and can
also serve as an indicator of employee happiness (Anger et al., 2021).

The process of human resources analytics has some steps listed as follows:
(a) Understanding the organization’s business goals, (b) Identifying the
metrics to be analyzed to achieve those goals, (c) Collecting and analyzing the
relevant data, (d) Obtaining insights into this data, and (e) Communicating
how this data affects the organization.
There are different industrial applications of human resources analytics.
Some of them have been summarized as follows:
General Electric (GE) experts collected different types of data relate to
working environments, i.e., from power plants to hospital equipment. GE's
analytics team used this data to make the machines working within it more
efficient. They stated that such business analyzes may increase productivity in
the US by 1.5% and increase the average national income by 30% over 20
years (Sağıroğlu and Koç, 2017).
FedEx, which has 155,000 employees worldwide, is one of the institutions
that have achieved success in human resources analytics. The company has
increased its use of analytics and has become able to access all the data needed.
“We're always looking at HR metrics.” says Bob Bennett, FedEx Executive
Vice President of Human Resources, who also said, “When evaluating our
ability to support business goals, we also look at business metrics”
(Hattangadi, 2019).
eQuest, a company that provides recruitment consultancy services to
financial companies, found in an analysis that 175,000 dollars had been spent
every year on 48 job websites, and 45 of them did not give timely feedback.
In addition, the company determined the words and phrases that candidates
use the most while searching and decided to change the titles and definitions
of job postings accordingly. As a result, candidate traffic increased by 175%
and the budget spent on job postings decreased by 50% (Doğan, 2019).

Conclusion

Human resources analytics not only activates processes in human resources

management and operations of the organization but also contributes to the
10 Cem Ufuk Baytar

development of the organization, its competitiveness, and brand identity as an

employer. Managing data correctly gives advantages, i.e., evaluating the level
of commitment of qualified manpower to the organization, analyzing the
intention to leave the job, and facilitating the management of talent by
estimating the potential of an employee. It should also be taken into account
that there is a need for concrete, measurable and interpretable information for
determining the right strategies. In this context, HR analytics plays an
important role in designing a strategic human resources approach by
(a) determining HR metrics in line with the priorities of the organization,
(b) tracking and evaluating these metrics periodically, and (c) strengthening
the talent pool by taking measures when necessary. Based on insights given in
this work, research in the future can be carried out to show the effects of
human resources metrics on the decisions of human resources management.

References

Anger, O., Tessema, M., Craft, J., Tsegai, S., 2021. A framework for assessing the
effectiveness of hr metrics and analytics: the case of an American healthcare
institution. Global Journal of Human Resource Management 9(1), 1-19.
Arai, K., Barakbah, A., 2007. Hierarchical k-means: An algorithm for centroids
initialization for k-means. Faculty of Science and Engineering, Saga University.
Balaban, M. E., Kartal, E., 2018. Veri madenciliği ve makine öğrenmesi temel algoritmaları
ve R dili ile uygulamaları. [Data mining and machine learning basic algorithms and
applications with R language.] İstabul: Çağlayan Publishing.
Berkhin, P., 2006. A survey of clustering data mining techniques in grouping
multidimensional data. Springer Berlin Heidelberg, 25-71.
Ceylan, Z., Gürsev, S., Bulkan, S., 2017. İki aşamalı kümeleme analizi ile bireysel
emeklilik sektöründe müşteri profilinin değerlendirilmesi. Bilişim Teknolojileri
Dergisi [Evaluation of customer profile in private pension sector with two-stage
cluster analysis. Journal of Information Technologies]10(4), 475-485.
Cömert, N., Kaymaz, M., 2019. Araç sigortası hilelerinde veri madenciliğinin kullanımı.
Marmara Üniversitesi, İktisadi ve İdari Bilimler Dergisi [The use of data mining in
car insurance cheats.] Marmara University, Journal of Economics and Administrative
Sciences] 41(2), 364-390.
Davenport, T. H., Harris, J.G., Morison, R., 2010. Analytics at Work: Smarter Decisions,
Better Results. Boston: Harvard Business Press.
Delen, D., Ram, S., 2018. Research Challenges and Opportunities in Business Analytics.
Journal of Business Analytics 1(1), 2-12.
Delen, D., Zolbanin, H.M., 2018. the Analytics Paradigm in Business Research. Journal of
Business Research 90, 186-195.
Data Analytics Applied to the Human Resources Industry 11

Doğan, G., 2019. Büyük Veri İnsan Kaynaklarına Ne Sunuyor? [What Does Big Data
Offer to Human Resources?] https://kaynakbaltas.com/kultur/buyuk-veri-insan-
kaynaklarina -ne-sunuyor/.
Dursun, B., 2013. İnsan Kaynakları Yönetimi. [Human Resources Management] Istanbul:
Beta publishing house.
Falletta, S.V., Combs, W. L., 2021. The HR analytics cycle: a seven-step process for
building evidence-based and ethical HR analytics capabilities. Journal of Work-
Applied Management 13(1), 51-68.
Fitz-Enz, J., Mattox, J., 2014. Predictive Analytics for Human Resources. New Jersey:
Wiley and SAS Business Series.
Gamgam, H., Altunkaynak, B., 2017. SPSS uygulamalı regresyon analizi [SPSS applied
regression analysis]. (2 ed.).
Han, J., Kamber, M., 2006. Data mining concepts and techniques. San Francisco: Elsevier,
Morgan Kaufmann Publishers.
Harrington, P., 2012. Machine learning in action. Manning Publications Co.
Hattangadi, V., 2019. Firms are turning to people analytics to achieve better efficiencies.
https://www.financialexpress.com/opinion/firms-are-turning-to-people-analytics-to-
achieve-better-efficiencies/1734443/
Heuvel, S., Bondarouk, T., 2016. The Rise (and Fall) of HR Analytics: A Study into the
Future Applications, Value, Structure, and System Support. Journal of Organizational
Effectiveness People and Performance 4(2), 127-148.
Jayanthi, R., 2020. A Study on Effectiveness of HR Metrics. International Journal of
Science and Research (IJSR) 9(2), 1630-1635.
Keçecioğlu, T., Oktay, S., 2010. İnsan Sermayesinin Sayısallaştırılmasının Dayanılmaz
Çekim Gücü [The Irresistible Attraction of the Digitization of Human Capital]. Dokuz
Eylül Üniversitesi, Sosyal Bilimler Enstitüsü Dergisi [Journal of Social Sciences
Institute] 12(3), 67-86.
Krishnamoorthi, S., Mathew, S.K., 2018. Business analytics and business value: A
comparative case study. Information & Management 55(5), 643-666.
Lalwani, P., 2022. What Is HR Analytics? Definition, Importance, Key Metrics, Data
Requirements, and Implementation. Accessed June 20. https://www.spiceworks.com/
hr/hr-analytics/articles/what-is-hr-analytics/.
Mortenson, M. J., Doherty, N. F., Robinson, S., 2015. Operational Research from
Taylorism to Terabytes: A Research Agenda for the Analytics Age. European Journal
of Operational Research, 241(3), 583-595.
Olavsrud, T., 2022. What is data analytics? Analyzing and managing data for decisions.
Accessed June 15. https://www.cio.com/article/191313/what-is-data-analytics-
analyzing-and-managing-data-for-decisions.html.
Özdemir, A., Sağlam, R., Bilen, B. B., 2018. Eğitim Sisteminde Veri Madenciliği
Uygulamaları ve Farkındalık Üzerine Bir Durum Çalışması. Atatürk Üniversitesi
Sosyal Bilimler Enstitüsü Dergisi [A Case Study on Data Mining Applications and
Awareness in Education System. Journal of Atatürk University Institute of Social
Sciences] 22 (Özel Sayı), 2159-2172.
Sağıroğlu, Ş., Koç, O., 2017. Büyük Veri ve Açık Veri Analitiği: Yöntemler ve
Uygulamalar. Ankara: Grafiker publishing house.
12 Cem Ufuk Baytar

Sayad, S., 2022. Decision tree regression. Accessed May 1 https://www.saedsayad.com/

decision_tree_reg.htm.
Stedman, C., 2022. Data analytics (DA). Accessed June 2 https://www.techtarget.com/
searchdatamanagement/definition/data-analytics.
Silahtaroğlu, G., 2013. Veri Madenciliği Kavram ve Algoritmaları. [Data Mining Concepts
and Algorithms.] Papatya publisher.
Steinbach, M., Karypis, G., Kumar, V., 2000. A comparison of document clustering
techniques. The sixth ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining.
Yıldırım, P., Birant, D., 2018. Bulut bilişimde veri madenciliği tekniklerinin uygulanması:
Bir literatür taraması. Pamukkale Üniv Muh Bilim Dergisi [Application of data mining
techniques in cloud computing: A literature review. Pamukkale University Journal of
Muh Science] 24(2), 336-343.
Chapter 2

Toxicogenomics Data Mining as a Promising

Prioritization Tool in Toxicity Testing

Katarina Živančević1,2,†, Dragica Bozic1,†,

Katarina Baralić1,† and Danijela Đukić-Ćosić1,*
1Department of Toxicology “Akademik Danilo Soldatović,”
University of Belgrade - Faculty of Pharmacy, Belgrade, Serbia
2Institute of Physiology and Biochemistry “Ivan Đaja,”

University of Belgrade - Faculty of Biology, Belgrade, Serbia

Abstract

Within regulatory agencies and toxicological research teams, there is an

increasing interest in developing, testing, and using novel techniques,
such as -omics, to analyse chemical risks even more efficiently.
Toxicogenomics explores the connections between genes and
environmental stress in disease aetiology by combining measurements of
families of biomolecules with bioinformatics and traditional toxicology.
It has a potential to predict gene functions within specific biological
pathways, identify genomic biomarkers and reduce gene sets.
Furthermore, it provides mixture evaluation techniques, taking into
account all the potential chemical, gene, protein, metabolite, and network
interactions that may be significant in triggering mixture toxicities
(antagonistic, additive/synergistic action). Hence, toxicogenomics data
mining might be viewed as an important stepping stone for further in
vitro and in vivo analyses, enabling the reduction of time and cost of the
overall analysis, as well as the number of laboratory animals, in
accordance with the animal welfare and the 3R principle. Having all of

*
Corresponding Author’s Email: [email protected].
†
The authors have equally contributed the research.

In: The Future of Data Mining

Editor: Cem Ufuk Baytar
ISBN: 979-8-88697-250-4
© 2022 Nova Science Publishers, Inc.
14 Katarina Živančević, Dragica Bozic, Katarina Baralić et al.

this in mind, in this chapter, toxicogenomics data mining will be

presented as a promising functional annotation-based prioritization tool,
with a concise summary of potentially useful software and tools,
including their major advantages and limitations.

Keywords: bioinformatics, database, software, tools, in silico

Introduction

Traditional toxicological approaches, such as animal research, are limited by

cost, time, and ethical considerations (Basile et al., 2019; Prior et al., 2019).
With the goal of reducing animal testing by 2035, the US Environmental
Protection Agency (EPA) proposed in silico modeling for assessing multiple
toxicity end-points of the tested chemicals (Kar and Leszczynski, 2020),
recommending the application of numerous computational techniques capable
of analyzing, simulating, visualizing, or predicting toxicity (Myatt et al.,
2018). In silico models are extremely complicated, relying on the information
obtained by statistics, mathematics, bioinformatics, biochemistry, and
molecular biology (Đukić-Ćosić et al., 2021; Shegokar, 2020). Furthermore,
massive amounts of toxicity data have been accumulated in numerous
databases during the last few decades. In most cases, they are used to access
experimental data in a convenient manner and contain experimental evidence
that may serve as a foundation for the development of new scientific
hypotheses (Helma et al., 2000). High-throughput toxicity screening tests are
considered capable of replacing many previously used in vivo tests, with non-
profit organizations, regulatory agencies (EPA, Health Canada, European
Chemicals Agency-ECHA), and international organizations (Organization for
Economic Cooperation and Development-OECD) advocating for this change
and promoting initiatives to reduce the amount of required animal-based tests
(Oki and Edwards, 2016). Toxicogenomic methods, which may include not
only genomics, but also transcriptomics, proteomics, metabolomics, and
epigenomics, seek to understand molecular causes of toxicity from a holistic
point of view. Most often, their additional goal is to identify expression
patterns (i.e., biomarkers) with aim to predict toxicity or genetic vulnerability
to toxic substances. This discipline has progressed as a result of improvements
in mechanistic toxicology, toxicity evaluations efficiency, chemical safety
assessments, and novel drug development tests (Koedrith et al., 2013). It has
a potential to add value in predicting gene functions within specific biological
Toxicogenomics Data Mining as a Promising Prioritization Tool … 15

pathways, as well as in reducing gene sets (Baralić et al., 2022). Furthermore,

it provides mixture evaluation techniques, taking into account all the potential
chemical, gene, protein, metabolite, and network interactions that may be
significant in triggering mixture toxicities (antagonistic, additive/synergistic
action) (Baralić et al., 2021a; Živančević et al., 2021). Hence, toxicogenomics
data mining might be viewed as an important stepping stone for further in vitro
and in vivo analysis, enabling the reduction of time and cost of the overall
analysis, as well as the number of laboratory animals, in accordance with the
animal welfare and the 3R principle (Bozic et al., 2022).
Having all of this in mind, in this chapter, toxicogenomics data mining
will be presented as a promising functional annotation-based prioritization
tool, with a concise summary of potentially useful software and tools,
including their major strength and limitations.

Useful Databases and Tools for Data Mining in Toxicology

The rapid advance in toxicogenomics has led to the development of distinct

computational tools that can facilitate the use of publicly available data for
toxicology research in compliance with the 3R principle. Here, we will
mention some of the most frequently used databases, together with the tools
for their analysis.
Comparative Toxicogenomic Database (CTD; http://ctdbase.org/) is a
public resource integrated with the data from OMIM, MeSH, NCBI Gene, GO,
KEGG and Reactome Pathway databases, developed with the aim to capture
the information on chemicals, genes and molecular mechanisms behind
chemically-induced diseases (Grondin et al., 2021). Firstly, CTD focused on
chemical-gene, chemical-disease, and gene-disease interactions. However,
over the last decade, it has evolved and developed, capturing new curation
paradigms such as chemical-exposure statements and chemical-phenotype
interactions (Davis et al., 2020). Currently, CTD contains more than 2.6
million of manually curated interactions for almost 17 000 chemicals, more
than 53 000 genes, and 7,267 diseases extracted from 139,720 carefully
revised references (http://ctdbase.org/about/dataStatus.go).
In the field of toxicology, the application of CTD has become a powerful
tool for predicting potentially hazardous effects of a given molecule or even
mixtures (Boverhof and Zacharewski, 2006). The aim of capturing
chemical(s) interacting genes is to further assess the ability of a given chemical
to induce or reduce the expression of protein-coding genes, protein synthesis
16 Katarina Živančević, Dragica Bozic, Katarina Baralić et al.

and/or secretion, which might suggest the ability of the investigated chemical
to disrupt different molecular pathways/processes in the organism (Cronin et
al., 2009; Živančević et al., 2019).
The in-depth understanding of the selected gene-data can be achieved by
the integration of the data from CTD with other mining tools/databases such
as ToppGene Suite, The Database for Annotation, Visualization and
Integrated Discovery (DAVID), DisGeNET or Cytoscape, open source
platform for visualization of complex networks and their integration with any
type of attribute data.
ToppGene Suite (https://toppgene.cchmc.org/) is a publicly available
website that can be used for gene list functional enrichment and candidate gene
prioritization. This is accomplished by using either functional annotations or
network analysis, identification and prioritization of novel candidate genes in
interactome, a protein-protein interaction network (Chen et al., 2009). This is
particularly useful for gaining the mechanistic knowledge of substance-
induced toxicity or adverse outcome. Similarly, DAVID (https://david.
ncifcrf.gov/) was developed with an idea to provide a comprehensive set of
functional annotation tools used for elucidating biological meaning of genes.
Users can easily add large gene-lists, change background populations, select
species and categories and reset functional parameters for data analysis.
Simultaneous application of all available tools during the analysis is also
possible, as well as formatting the final output (Jiao et al., 2012). In other
words, DAVID facilitates the transition from genome-scale datasets collection
to biological meaning (Dennis et al., 2003).
Another useful resource designed as the largest publicly available
collection of genes and variants associated to human diseases is DisGeNET
(https://www.disgenet.org/). The current version of DisGeNET (v7.0)
contains 1,134,942 gene-disease associations, curated for 21,671 genes and
30,170 diseases, disorders, traits, and clinical or abnormal human phenotypes.
Moreover, 369,554 variant-disease associations between almost 195,000
variants and more than 14,000 diseases, traits, and phenotypes are described
in the latest version (Piñero et al., 2020). Thus, combining CTD data-mining
with DisGeNET can be viewed as a useful way for predicting chemically-
induced abnormal human phenotypes (Baralić et al., (2022)).
Finally, one of the most famous software used for the genome-related
biological research is Cytoscape - a network visualization tool integrated with
different plug-ins (Shannon et al., 2003). For example, GeneMANIA plug-in
can be used for identification of the genes most related to a query gene set by
using a guilt-by-association approach (Montojo et al., 2014), while
Toxicogenomics Data Mining as a Promising Prioritization Tool … 17

CytoHubba uses several topological algorithms to explore the most important

nodes/hubs and fragile motifs in the generated network. The advantage of this
plug-in is to prioritize and predict essential genes/proteins in complex
biological networks (Chin et al., 2014). For example, Uddin and Wang (2022)
were able to discover the key tumor-stroma associated biomarkers which
correlated with dysregulated pathways, tumor immunity, tumor progression,
and clinical outcomes in breast cancer by exploring the hub genes, oncogenes
and protein kinases (Uddin et al., 2020). On the other hand, ClueGo is used
for single and comparison cluster analysis that can be considered helpful for
identification of pathways regulated by gene sets obtained by in silico and/or
experimental analyses. ClueGo is updated with the newest files from Gene
Ontology, KEGG, WikiPathways and Reactome. Its functionality is improved
by CluePedia, which calculates linear and non-linear statistical dependencies
of genes, proteins and miRNAs integrated into a network with ClueGO
terms/pathways (Bindea et al., 2013).

Data Mining Examples

As an example, Davis et al., (2008) presented data mining using arsenic as an

example. They demonstrated how CTD can be applied as an insight into the
biological functions and molecular networks affected by the exposure to
arsenic, including stress response, apoptosis, cell cycle, and specific protein
signaling pathways. Integrating arsenic-gene data with gene-disease data
provided a list of diseases connected to arsenic exposure, as well as a list of
genes that may explain this association (Davis et al., 2008).
Other studies also used this in silico approach to explore metal-induced
toxic effects. In a study conducted by Živančević et al., (2021), CTD and
Cytoscape software were utilized to examine the impact of environmentally
relevant toxic metals (lead, methylmercury (organic form of mercury),
cadmium, and arsenic) on molecular pathways implicated in the development
of amyotrophic lateral sclerosis (ALS), Parkinson’s disease (PD), and
Alzheimer’s disease (AD). The findings revealed 7, 13, and 14 genes common
to all the tested metals and related to ALS, PD, and AD, respectively, while
SOD2 gene was highlighted as the most important mutual gene for all the
selected diseases. The primary identified biochemical pathways involved in
the development of the investigated neurodegenerative disorders were
oxidative stress, folate metabolism, vitamin B12, AGE-RAGE, and apoptosis.
The examined metal mixture was also found to influence various biological
18 Katarina Živančević, Dragica Bozic, Katarina Baralić et al.

processes. Glutathione metabolic process was listed as the most important for
ALS, cellular response to toxic substance for PD, and neuron death for AD
(Živančević et al., 2021).
In order to develop a feasible set of genomic biomarkers, Baralić et al.,
(2022) used in silico toxicogenomic data mining to investigate the link
between phthalates and bisphenol A (BPA) co-exposure and obesity, as well
as its comorbid conditions. For this purpose, CTD database was used as the
main in silico tool, along with GeneMania, ToppGene Suite and DisGeNET.
As a result, 7 mutual genes (6 relevant to humans: CCL2, IL6, LPL, PPARG,
SERPINE1, and TNF) were identified in all the investigated obesity
comorbidities. Additionally, of all the extracted genes, PPARG and LPL were
found most closely linked to obesity (Baralić et al., 2022). The same group of
authors investigated the connections between DEHP, DBP and BPA co-
exposure and type 2 diabetes mellitus (T2DM), male infertility and asthma
(Baralić et al., 2021a, 2021b, 2021c). In the aforementioned studies, CTD and
ToppGene Suite were used as the main data mining tools, along with the
Cytoscape software. The analysis revealed 44, 20 and 24 genes common to all
the investigated substances and involved in the development of T2DM, male
infertility and asthma, respectively. Apoptosis and oxidative stress were
singled out as the most important mechanisms of both DMT2 and asthma
(Baralić et al., 2021a, 2021c), while inflammation was additionally
highlighted as a particularly important pathway in asthma development
(Baralić et al., 2021a). Proposed mechanisms of DEHP, DBP and BPA
mixture induced testicular toxicity were effects on sex differentiation,
metabolism, nuclear receptors, aryl hydrocarbon receptors, apoptosis, and
oxidative stress (Baralić et al., 2021b). Since oxidative stress was highlighted
as one of the main mechanisms of phthalate and BPA mixture for of all of
these disorders, data mining results were further validated by additional in vivo
experiments, further suggesting additive effects of the investigated substances
(Baralić et al., 2021a, 2021b, 2021c).
Another study which combined in silico with in vivo results was
performed by Dong et al., (2018). As a follow-up to the in vivo zebrafish
experiment, these authors performed CTD data mining for BPA and DBP. In
their study, by conducting CTD analysis, they discovered 4826 and 14737
interactions with various genes/proteins for DBP and BPA, respectively.
When orthologous human genes in BPA- and DBP-treated embryos were
compared to curated BPA- and DBP-interacting proteins in CTD, 18 and 13
mutual proteins were found, respectively, while 7 and 9 of the 18 and 13
proteins were part of the BPA- and DBP-predicted pathways, respectively.
Toxicogenomics Data Mining as a Promising Prioritization Tool … 19

The results of CTD and proteomic analysis further highlighted the possible
impacts of BPA and DBP on the hypothesized networks (Dong et al., 2018).
Furthermore, Baralić et al., (2020) assessed both risks and benefits of the
COVID-19 treatment with candidate drug combinations: lopinavir/ritonavir
and chloroquine/hydroxychloroquine + azithromycin, by using CTD,
Cytoscape software and ToppGene Suite portal as a foundation in their
research. Lopinavir/ritonavir increased the expression of specific genes
involved in the immune response and lipid metabolism (IL6, ICAM1, CCL2,
TNF, APOA1, etc.). Chloroquine/hydroxychloroquine + azithromycin
interacted with six genes (CCL2, CTSB, CXCL8, IL1B, IL6 and TNF),
whereas chloroquine + azithromycin acted on two extra genes (BCL2L1 and
CYP3A4) (Baralić et al., 2020).
Using the immunomodulator sulforaphane (SFN) as a case study, Bozic
et al., (2021) applied toxicogenomic data mining (CTD, ToppGene Suite
portal and Reactome Knowledgebase) to investigate molecular mechanisms
and pathways which might be targeted in cancer treatment. Sulforaphane
interacted with 1896 different proteins, the most important being NFE2L2,
NQO1, HMOX1, GCLC, TXNRD1, IL1B, IFNG, AGT, KEAP1, and CASP3.
A direct evidence was found in CTD that SFN interacted with a total of 169
genes connected to its therapeutic effect against various forms of cancer,
including hepatocellular carcinoma, colorectal neoplasms, uterine cervical
neoplasms, and adenomatous polyposis coli (Bozic et al., 2021). The same
group of authors further explored SFN induced adverse effects in colorectal
carcinoma patients. In this study, gene network analysis was performed using
Cytoscape plug-in STRING. Functional annotation clustering was performed
by the Cytoscape plug-in BINGO, while SFN interacting genes were obtained
from CTD. SFN was found to increase the expression of TIMP1, AURKA,
and CEP55, and decrease the expression of CRYAB, PLCE1, and MMP28,
which may contribute to the advancement of colorectal carcinoma (CTD).
According to the pathway enrichment analyses, SFN enhanced RUNX2 and
AURKA activation through TPX2 and IL-10 signaling (Bozic et al., 2022).

Advantages

Conventional toxicity testing and human health risk assessment strategies

offer limited insight into the fundamental molecular mechanisms that
contribute to adverse outcomes (Chepelev et al., 2015). In order to improve
and facilitate risk assessment of thousands of chemicals with insufficient
20 Katarina Živančević, Dragica Bozic, Katarina Baralić et al.

toxicity data, significant adjustment of the conventional approach is required

(Chepelev et al., 2015). Since toxicogenomics explores the connections
between genes and environmental stress in disease aetiology by combining
measurements of families of biomolecules with bioinformatics and traditional
toxicology, it has a potential to add value in predicting gene functions within
specific biological pathways, as well as to identify genomic biomarkers and
reduce gene sets (Boverhof and Zacharewski, 2006; Breda et al., 2014; Liu et
al., 2020; Tung et al., 2020). In the light of this, toxicogenomics biomarkers
are of particular importance having in mind that they can be discovered earlier
than histopathological or clinical phenotypes (Ulrich and Friend, 2002).
Furthermore, toxicogenomic approach enables high-throughput screening of
toxic substances, which could lead to facilitated prioritization of chemicals in
toxicity testing (Franzosa et al., 2021). Moreover, combining toxicogenomics
with other bioinformatics approaches, such as cheminformatics, could
facilitate rapid prediction of the pharmacokinetic and physicochemical key
parameters for examined molecules and, thus, provide integrated data sets
which could be used in regulatory purposes (Daina et al., 2017).
Toxicogenomics approaches could also facilitate adverse outcome pathways
(AOP) development and enhance read-across strategies (Liu et al., 2020).
AOP is a conceptual framework which connects molecular initiating events
and key events with outcomes and adverse effects in risk assessment (Sewell
et al., 2018). It integrates molecules, gene activities, and causal adverse
outcomes. In this context, toxicogenomics represents an important resource
for linking the information among the AOP components (key events and their
relationships) (Liu et al., 2020). On the other hand, read-across assumes that
the two ‘similar’ compounds are likely to share a similar toxicity profile based
on their structural similarities (Liu et al., 2020). However, if the two chemicals
have similar biological profiles, they might also share an adverse outcome (Liu
et al., 2020). Thus, diverse biological data, including toxicogenomic, could
also be used for biological fingerprint profiling of chemicals. Thus,
biologically-based read-across could be viewed as a complement to the
structure-based read-across (Ball et al., 2016). Toxicogenomic data also offers
strategies to derive molecular points of departure (POD) for the dose-response
assessment (Phillips et al., 2015). Exposure to mixtures of toxic substances
that are not entirely known is a plausible scenario which should be explored
in occupational or environmental toxicology (Martins et al., 2019).
Conveniently, toxicogenomics also provides mixture evaluation techniques,
taking into consideraton all the potential chemical, gene, protein, metabolite,
and network interactions that may be important in triggering mixture toxicities
Toxicogenomics Data Mining as a Promising Prioritization Tool … 21

(Altenburger et al., 2012). Furthermore, this discipline could help to

distinguish the concept of interactions between toxic substances (antagonistic,
additive/synergistic) (Martins et al., 2019). Finally, it allows the prediction of
potential long-term effects of exposures, enabling the reduction of time and
cost of the overall analysis, as well as the number of laboratory animals, in
accordance with the animal welfare and the 3R principle (Yauk et al., 2019).
Hence, toxicogenomics data mining might be viewed as an important stepping
stone for further in vitro, in vivo and human biomonitoring research.

Limitations

Although toxicogenomics approach is a promising functional annotation-

based prioritization tool, its major limitations should not be overlooked.
Online sources (i.e., CTD, GeneMANIA plug-in, ToppGene Suite portal, etc.)
rely on limited available data and, therefore, the quality of the analysis
depends on the data relevance (Chen et al., 2009; Živančević et al., 2021). Due
to the fact that conclusions based on toxicogenomic data mining imply
drawing statistical associations between chemical-gene-disease relationships,
it is not possible to take into consideration some important aspects, such as
dose-response relationship, route and duration of exposure to chemicals, along
with the individual sensitivity of exposed subjects (Davis et al., 2009). The
main challenge of utilization of toxicogenomic data for risk assessment
applications is the lack of internationally harmonized rules for toxicogenomic
experimental methods, quality standards, references, and analytical
frameworks that provide the criteria and standards required for global
consistency (Vachon et al., 2017; Yauk et al., 2019). Additionally, the lack of
internationally approved strategies or frameworks for applying
toxicogenomics in specific risk assessment, as well as the lack of expertise and
training in toxicogenomics within the regulatory community in certain areas,
represent a challenge in applying toxicogenomics data in the regulatory
purposes (Pain et al., 2020; Yauk et al., 2019). Inadequate regulatory
capability to evaluate provided toxicogenomic data, as well as limited
applicability to certain areas are also issues which need to be addressed and
resolved (Yauk et al., 2019). Incomplete validation of pathway alterations in
specific diseases, or incomplete validation of the observed changes in gene
expression are proportional to the severity of the adverse effect (Yauk et al.,
2019). Also, the lack of high-quality toxicogenomic data make its integration
in toxicity testing impossible (Krewski et al., 2020; Yauk et al., 2019).
22 Katarina Živančević, Dragica Bozic, Katarina Baralić et al.

Some of the challenges in the mixture toxicity analysis are that detected
signals must be linked to the predefined response chains. It is also necessary
to understand the crosstalk and convergence of the pathways, considering that
joint responses might switch between independent and concentration additive,
or non-interactive and interactive. Since quantitative mixture assessment
typically relies on monotonous changes in response represented in sigmoidal
concentration response relationships, U-shaped curves present specific
challenges (Altenburger et al., 2012). Elucidation of the sequences of
biological responses, as well as separation of toxicokinetic from
toxicodynamic responses could also represent a challenge in interpretation of
toxicogenomic data (Altenburger et al., 2012). “Omics” techniques, are
generally complex to compute and understand. Thus, toxicogenomics data
mining could be demanding, especially in the integration of different data
formats (Martins et al., 2019), especially considering the variability of
available data and measured endpoints (Liu et al., 2020). The reproducibility
of toxicogenomic data, including both biological and technical sides, could
also make drawing conclusions difficult (Liu et al., 2020). Applicability
domain of toxicogenomic data is also an important factor which may limit the
analysis. A false positive result could be generated by biologically-based read-
across, which relies on toxicogenomic data (Liu et al., 2020). The fast
development of toxicogenomic approach has greatly expanded its impact on
toxicity data testing. However, aforementioned challenges should be
addressed before toxicogenomic data mining can meet its full potential as a
promising prioritization tool in toxicity testing (Liu et al., 2020).

Conclusion

The use of bioinformatic data mining methods in toxicology is growing, and

various in silico toxicity prediction tools and databases are updated on a
regular basis to meet the customer demands. In this chapter, toxicogenomics
data mining was marked as a prospective functional annotation-based
prioritizing approach, while its primary advantages and limitations were
addressed. All things considered, toxicogenomics data mining might be seen
as a useful tool for predicting and providing better understanding of
complicated mechanisms of toxicity, identifying genomic biomarkers,
reducing gene sets and evaluating mixtures of toxic substances. Finally, it can
be viewed as a valuable prioritization tool for further laboratory investigations,
Toxicogenomics Data Mining as a Promising Prioritization Tool … 23

allowing a decrease in total time and expense, as well as the number of

laboratory animals.

Acknowledgments

Ministry of Education, Science and Technological Development of the

Republic of Serbia in the framework of scientific cooperation with the
People’s Republic of China (451-03-1203/2021-09).

References

Altenburger, R., Scholz, S., Schmitt-Jansen, M., Busch, W., Escher, B. I., 2012. Mixture
toxicity revisited from a toxicogenomic perspective. Environmental Science and
Technology 46, 2508-2522. https://doi.org/10.1021/es2038036.
Ball, N., Cronin, M. T. D., Shen, J., Blackburn, K., Booth, E. D., Bouhifd, M., Donley, E.,
Egnash, L., Hastings, C., Juberg, D. R., Maertens, A., Marty, S., Naciff, J. M., Palmer,
J., Pamies, D., 2016. Toward Good Read-Across Practice (GRAP) Guidance. Altex 33,
149-166.
Baralić, K., Bozic, D., Živančević, K., Milenković, M., Javorac, D., Marić, Đ., Miljaković,
E. A., Djordjevic, A. B., Vukomanović, P., Ćurčić, M., Bulat, Z., Antonijević, B.,
Đukić-Ćosić, D., 2021a. Integrating in silico with in vivo approach to investigate
phthalate and bisphenol A mixture-linked asthma development: Positive probiotic
intervention. Food and Chemical Toxicology 158, 112671. https://doi.org/10.1016/
j.fct.2021.112671.
Baralić, K., Jorgovanović, D., Živančević, K., Antonijević Miljaković, E., Antonijević, B.,
Buha Djordjevic, A., Ćurčić, M., Đukić-Ćosić, D., 2020. Safety assessment of drug
combinations used in COVID-19 treatment: in silico toxicogenomic data-mining
approach. Toxicology and Applied Pharmacology 406, 115237. https://doi.org/10.
1016/j.taap.2020.115237.
Baralić, K., Jorgovanović, D., Živančević, K., Buha Djordjević, A., Antonijević
Miljaković, E., Miljković, M., Kotur-Stevuljević, J., Antonijević, B., Đukić-Ćosić, D.,
2021b. Combining in vivo pathohistological and redox status analysis with in silico
toxicogenomic study to explore the phthalates and bisphenol A mixture-induced
testicular toxicity. Chemosphere 267, 129296. https://doi.org/10.1016/j.chemosphere.
2020.129296.
Baralić, K., Živančević, K., Božić, D., Jennen, D., Buha Djordjevic, A., Antonijević
Miljaković, E., Đukić-Ćosić, D., 2022. Potential genomic biomarkers of obesity and
its comorbidities for phthalates and bisphenol A mixture: In silico toxicogenomic
approach. Biocell 46, 519-533. https://doi.org/10.32604/biocell.2022.018271.
Baralić, K., Živančević, K., Jorgovanović, D., Javorac, D., Radovanović, J., Gojković, T.,
Djordjevic, A. B., Ćurčić, M., Mandinić, Z., Bulat, Z., Antonijević, B., Đukić-Ćosić,
24 Katarina Živančević, Dragica Bozic, Katarina Baralić et al.

D., 2021c. Probiotic reduced the impact of phthalates and bisphenol A mixture on type
2 diabetes mellitus development: merging bioinformatics with in vivo analysis. Food
and Chemical Toxicology 154, 112325. https://doi.org/10.1016/j.fct.2021.112325.
Basile, A. O., Yahi, A., Tatonetti, N. P., 2019. Artificial Intelligence for Drug Toxicity and
Safety. Trends in Pharmacological Sciences 40, 624-635. https://doi.org/10.1016/
j.tips.2019.07.005.
Bindea, G., Galon, J., Mlecnik, B., 2013. CluePedia Cytoscape plugin: Pathway insights
using integrated experimental and in silico data. Bioinformatics 29, 661-663.
https://doi.org/10.1093/bioinformatics/btt019.
Boverhof, D. R., Zacharewski, T. R., 2006. Toxicogenomics in risk assessment:
Applications and needs. Toxicological Sciences 89, 352-360. https://doi.org/10.1093/
toxsci/kfj018.
Bozic, D., Baralić, K., Živančević, K., Miljaković, E. A., Ćurčić, M., Antonijević, B.,
Djordjević, A. B., Bulat, Z., Zhang, Y., Yang, L., Đukić-Ćosić, D., 2022. Predicting
sulforaphane-induced adverse effects in colon cancer patients via in silico
investigation. Biomedicine & Pharmacotherapy 146, 112598. https://doi.org/10.1016/
j.biopha.2021.112598.
Bozic, D., Živančević, K., Baralić, K., Javorac, D., 2021. Applying in silico toxicogenomic
data mining to predict molecular mechanisms and pathways against carcinoma:
immunomodulator sulforaphane as a case study, in: 1st International Conference on
Chemo and Bioinformatics, October 26-27, Kragujevac, Serbia. pp. 470-473.
Breda, S. G. J. Van, Claessen, S. M. H., Lo, K., Herwijnen, M. Van, Gaj, S., Kok, T. M. C.
M. De, Kleinjans, J. C. S., 2014. Epigenetic mechanisms underlying arsenic ‑
associated lung carcinogenesis. Archives of Toxicology 89, 1959-1969. https://doi.org/
10.1007/s00204-014-1351-2.
Chen, J., Bardes, E. E., Aronow, B. J., Jegga, A. G., 2009. ToppGene Suite for gene list
enrichment analysis and candidate gene prioritization. Nucleic Acids Research 37,
305-311. https://doi.org/10.1093/nar/gkp427.
Chepelev, N. L., Moffat, I. D., Labib, S., Bourdon-Lacombe, J., Kuo, B., Buick, J. K.,
Lemieux, F., Malik, A. I., Halappanavar, S., Williams, A., Yauk, C. L., 2015.
Integrating toxicogenomics into human health risk assessment: Lessons learned from
the benzo[a]pyrene case study. Critical Reviews in Toxicology 45, 44-52. https://doi.
org/10.3109/10408444.2014.973935.
Chin, C. H., Chen, S. H., Wu, H. H., Ho, C. W., Ko, M. T., Lin, C. Y., 2014. cytoHubba:
Identifying hub objects and sub-networks from complex interactome. BMC Systems
Biology 8, S11. https://doi.org/10.1186/1752-0509-8-S4-S11.
Cronin, M. T. D., Bajot, F., Enoch, S. J., Madden, J. C., Roberts, D. W., Schwöbel, J., 2009.
The in chemico-in silico interface: Challenges for integrating experimental and
computational chemistry to identify toxicity. ATLA Alternatives to Laboratory
Animals 37, 513-521. https://doi.org/10.1177/026119290903700508.
Daina, A., Michielin, O., Zoete, V., 2017. SwissADME: A free web tool to evaluate
pharmacokinetics, drug-likeness and medicinal chemistry friendliness of small
molecules. Scientific Reports 7, 1-13. https://doi.org/10.1038/srep42717.
Davis, A. P., Murphy, C. G., Rosenstein, M. C., Wiegers, T. C., Mattingly, C. J., 2008. The
Comparative Toxicogenomics Database facilitates identification and understanding of
Toxicogenomics Data Mining as a Promising Prioritization Tool … 25

chemical-gene-disease associations: arsenic as a case study. BMC Medical Genomics

1, 1-12. https://doi.org/10.1186/1755-8794-1-48.
Davis, A. P., Murphy, C. G., Saraceni-Richards, C. A., Rosenstein, M. C., Wiegers, T. C.,
Mattingly, C. J., 2009. Comparative Toxicogenomics Database: A knowledgebase and
discovery tool for chemical-gene-disease networks. Nucleic Acids Research 37, 786-
792. https://doi.org/10.1093/nar/gkn580.
Davis, A. P., Wiegers, T. C., Grondin, C. J., Johnson, R. J., Sciaky, D., Wiegers, J.,
Mattingly, C. J., 2020. Leveraging the comparative toxicogenomics database to fill in
knowledge gaps for environmental health: A test case for air pollution-induced
cardiovascular disease. Toxicological Sciences 177, 392-404. https://doi.org/10.1093/
toxsci/kfaa113.
Dennis, G., Sherman, B. T., Hosack, D. A., Yang, J., Gao, W., Lane, H. C., Lempicki, R.
A., 2003. DAVID: Database for Annotation, Visualization, and Integrated Discovery.
Genome biology 4. https://doi.org/10.1186/gb-2003-4-9-r60.
Dong, X., Qiu, X., Meng, S., Xu, H., Wu, X., Yang, M., 2018. Proteomic profile and
toxicity pathway analysis in zebra fi sh embryos exposed to bisphenol A and di- n -
butyl phthalate at environmentally relevant levels. Chemosphere 193, 313-320.
https://doi.org/10.1016/j.chemosphere.2017.11.042.
Đukić-Ćosić, D., Baralić, K., Jorgovanović, D., Živančević, K., Javorac, D., Stojilković,
N., Radović, B., Marić, Đ., Ćurčić, M., Buha-Đorđević, A., Bulat, Z., Antonijević-
Miljaković, E., Antonijević, B., 2021. “In silico” toxicology methods in drug safety
assessment. Arhiv za farmaciju 71, 257-278. https://doi.org/10.5937/arhfarm71-
32966.
Franzosa, J. A., Bonzo, J. A., Jack, J., Baker, N. C., Kothiya, P., Witek, R. P., Hurban, P.,
Siferd, S., Hester, S., Shah, I., Ferguson, S. S., Houck, K. A., Wambaugh, J. F., 2021.
High-throughput toxicogenomic screening of chemicals in the environment using
metabolically competent hepatic cell cultures. npj Systems Biology and Applications
7. https://doi.org/10.1038/s41540-020-00166-2.
Grondin, C. J., Davis, A. P., Wiegers, J. A., Wiegers, T. C., Sciaky, D., Johnson, R. J.,
Mattingly, C. J., 2021. Predicting molecular mechanisms, pathways, and health
outcomes induced by Juul e-cigarette aerosol chemicals using the Comparative
Toxicogenomics Database. Current Research in Toxicology 2, 272-281. https://doi.
org/10.1016/j.crtox.2021.08.001.
Helma, C., Gottmann, E., Kramer, S., 2000. Knowledge discovery and data mining in
toxicology. Statistical Methods in Medical Research 9, 329-358. https://doi.org/10.
1177/096228020000900403.
Jiao, X., Sherman, B. T., Huang, D. W., Stephens, R., Baseler, M. W., Lane, H. C.,
Lempicki, R. A., 2012. DAVID-WS: A stateful web service to facilitate gene/protein
list analysis. Bioinformatics 28, 1805-1806. https://doi.org/10.1093/bioinformatics/
bts251.
Kar, S., Leszczynski, J., 2020. Open access in silico tools to predict the ADMET profiling
of drug candidates. Expert Opinion on Drug Discovery 15, 1473-1487. https://doi.
org/10.1080/17460441.2020.1798926.
Koedrith, P., Kim, H. L., Weon, J. Il, Seo, Y. R., 2013. Toxicogenomic approaches for
understanding molecular mechanisms of heavy metal mutagenicity and
26 Katarina Živančević, Dragica Bozic, Katarina Baralić et al.

carcinogenicity. International Journal of Hygiene and Environmental Health 216,

587-598. https://doi.org/10.1016/j.ijheh.2013.02.010.
Krewski, D., Andersen, M. E., Tyshenko, M. G., Krishnan, K., Hartung, T., Boekelheide,
K., Wambaugh, J. F., Jones, D., Whelan, M., Thomas, R., Yauk, C., Barton-Maclaren,
T., Cote, I., 2020. Toxicity testing in the 21st century: progress in the past decade and
future perspectives, Archives of Toxicology. Springer Berlin Heidelberg. https://doi.
org/10.1007/s00204-019-02613-4.
Liu, Z., Huang, R., Roberts, R., Tong, W., 2020. Toxicogenomics : A 2020 Vision. Trends
in Pharmacological Sciences 40, 92-103. https://doi.org/10.1016/j.tips.2018.12.001.
Martins, C., Dreij, K., Costa, P. M., 2019. The state-of-the art of environmental
toxicogenomics: Challenges and perspectives of “omics” approaches directed to
toxicant mixtures. International Journal of Environmental Research and Public
Health 16, 1-16. https://doi.org/10.3390/ijerph16234718.
Montojo, J., Zuberi, K., Rodriguez, H., Bader, G. D., Morris, Q., 2014. GeneMANIA: Fast
gene network construction and function prediction for Cytoscape. F1000Research 3,
1-7. https://doi.org/10.12688/f1000research.4572.1.
Myatt, G. J., Ahlberg, E., Akahori, Y., Allen, D., Amberg, A., Anger, L. T., Aptula, A.,
Auerbach, S., Beilke, L., Bellion, P., Benigni, R., Bercu, J., Booth, E. D., Bower, D.,
Brigo, A., Burden, N., Cammerer, Z., Cronin, M. T. D., Cross, K. P., Custer, L.,
Dettwiler, M., Dobo, K., Ford, K. A., Fortin, M. C., Gad-McDonald, S. E., Gellatly,
N., Gervais, V., Glover, K. P., Glowienke, S., Van Gompel, J., Gutsell, S., Hardy, B.,
Harvey, J. S., Hillegass, J., Honma, M., Hsieh, J. H., Hsu, C. W., Hughes, K., Johnson,
C., Jolly, R., Jones, D., Kemper, R., Kenyon, M. O., Kim, M. T., Kruhlak, N. L.,
Kulkarni, S. A., Kümmerer, K., Leavitt, P., Majer, B., Masten, S., Miller, S., Moser,
J., Mumtaz, M., Muster, W., Neilson, L., Oprea, T. I., Patlewicz, G., Paulino, A., Lo
Piparo, E., Powley, M., Quigley, D. P., Reddy, M. V., Richarz, A. N., Ruiz, P.,
Schilter, B., Serafimova, R., Simpson, W., Stavitskaya, L., Stidl, R., Suarez-
Rodriguez, D., Szabo, D. T., Teasdale, A., Trejo-Martin, A., Valentin, J. P., Vuorinen,
A., Wall, B. A., Watts, P., White, A. T., Wichard, J., Witt, K. L., Woolley, A.,
Woolley, D., Zwickl, C., Hasselgren, C., 2018. In silico toxicology protocols.
Regulatory Toxicology and Pharmacology 96, 1-17. https://doi.org/10.1016/j.yrtph.
2018.04.014.
Oki, N. O., Edwards, S. W., 2016. An integrative data mining approach to identifying
adverse outcome pathway signatures. Toxicology 350-352, 49-61. https://doi.org/10.
1016/j.tox.2016.04.004.
Pain, G., Hickey, G., Mondou, M., Crump, D., Hecker, M., Basu, N., Maguire, S., 2020.
Drivers of and obstacles to the adoption of toxicogenomics for chemical risk
assessment: Insights from social science perspectives. Environmental Health
Perspectives 128, 1-12. .https://doi.org/10.1289/EHP6500.
Phillips, J. R., Svoboda, D. L., Tandon, A., Patel, S., Mav, D., Kuo, B., Yauk, C. L., Yang,
L., Thomas, R. S., Gift, J. S., Davis, J. A., Olysyzk, L., Alex, B., Paules, R. S., Parham,
F., Saddler, T., Ruchir, R., Auerbach, S. S., 2015. BMDExpress 2: Enhanced
transcriptomic dose- response analysis workflow. Bioinformatics 14, 1780-2.
https://doi.org/10.1093/bioinformatics/xxxxx.
Toxicogenomics Data Mining as a Promising Prioritization Tool … 27

Piñero, J., Ramírez-Anguita, J. M., Saüch-Pitarch, J., Ronzano, F., Centeno, E., Sanz, F.,
Furlong, L. I., 2020. The DisGeNET knowledge platform for disease genomics: 2019
update. Nucleic Acids Research 48, D845-D855. https://doi.org/10.1093/nar/gkz1021.
Prior, H., Casey, W., Kimber, I., Whelan, M., Sewell, F., 2019. Reflections on the progress
towards non-animal methods for acute toxicity testing of chemicals. Regulatory
Toxicology and Pharmacology 102, 30-33. https://doi.org/10.1016/j.yrtph.2018.12.
008.
Sewell, F., Gellatly, N., Beaumont, M., Burden, N., Currie, R., de Haan, L., Hutchinson, T.
H., Jacobs, M., Mahony, C., Malcomber, I., Mehta, J., Whale, G., Kimber, I., 2018.
The future trajectory of adverse outcome pathways: a commentary. Archives of
Toxicology 92, 1657-1661. https://doi.org/10.1007/s00204-018-2183-2.
Shannon, P., Markiel, A., Ozier, O., Shannon, P., Markiel, A., Ozier, O., Baliga, N. S.,
Wang, J. T., Ramage, D., Amin, N., Schwikowski, B., Ideker, T., 2003. Cytoscape : A
Software Environment for Integrated Models of Biomolecular Interaction Networks
2498-2504. https://doi.org/10.1101/gr.1239303.
Shegokar, R., 2020. Preclinical testing-Understanding the basics first, Drug Delivery
Aspects. Elsevier Inc. https://doi.org/10.1016/b978-0-12-821222-6.00002-6.
Tung, C. W., Jen, H., Chia, C., Wang, C., Shan, S., Pinpin, W., 2020. Leveraging
complementary computational models for prioritizing chemicals of developmental
and reproductive toxicity concern : an example of food contact materials. Archives of
Toxicology 1-10. https://doi.org/10.1007/s00204-019-02641-0.
Uddin, M., Mustafa, F., Rizvi, T. A., Loney, T., Suwaidi, H. Al, Al-Marzouqi, A. H. H.,
Eldin, A. K., Alsabeeha, N., Adrian, T. E., Stefanini, C., Nowotny, N., Alsheikh-Ali,
A., Senok, A. C., 2020. SARS-CoV-2/COVID-19: Viral Genomics, Epidemiology,
Vaccines, and Therapeutic Interventions. Viruses 12, 526. https://doi.org/10.4324/978
1351118422-5.
Ulrich, R., Friend, S. H., 2002. Toxicogenomics and drug discovery: Will new technologies
help us produce better drugs? Nature Reviews Drug Discovery 1, 84-88. https://doi.
org/10.1038/nrd710.
Vachon, J., Campagna, C., Rodriguez, M. J., Sirard, M. A., Levallois, P., 2017. Barriers to
the use of toxicogenomics data in human health risk assessment: A survey of Canadian
risk assessors. Regulatory Toxicology and Pharmacology 85, 119-123. https://doi.org/
10.1016/j.yrtph.2017.01.008.
Yauk, C. L., Cheung, C., Barton-Maclaren, T. S., Boucher, S., Bourdon-Lacombe, J.,
Chauhan, V., Gagné, M., Gillespie, Z., Halappanavar, S., Honeyman, M., Jones, S. R.,
Jones-McLean, E., Labib, S., MacAulay, J., Moore, J., Paquette, M., Petronella, N.,
Semalulu, S., Slot, A., Vespa, A., Woodland, C. L. A., 2019. Toxicogenomic
applications in risk assessment at Health Canada. Current Opinion in Toxicology 18,
34-45. https://doi.org/10.1016/j.cotox.2019.02.005.
Živančević, K., Baralić, K., Jorgovanović, D., Buha Djordjević, A., Ćurčić, M., Antonijević
Miljaković, E., Antonijević, B., Bulat, Z., Đukić-Ćosić, D., 2021. Elucidating the
influence of environmentally relevant toxic metal mixture on molecular mechanisms
involved in the development of neurodegenerative diseases: In silico toxicogenomic
data-mining. Environmental Research 194. https://doi.org/10.1016/j.envres.2021.
110727.
28 Katarina Živančević, Dragica Bozic, Katarina Baralić et al.

Živančević, K., Baralić, K., Jorgovanović, D., Đukić-Ćosić, D., 2019. The Comparative
Toxicogenomics database: the influence of environmental chemicals on genes (In
Serbian). MD-Medical Data 11(3-4):159-164.
Chapter 3

Applications of Data Mining Algorithms

for Customer Recommendations
in Retail Marketing

Elif Delice1,*, Lütviye Özge Polatlı2,

İrem Düzdar Argun3 and Hakan Tozan4
1Management Information Systems, Istanbul Topkapı University, Istanbul, Turkey
2Healthcare Systems Engineering, Istanbul Medipol University, Istanbul, Turkey
3Industrial Engineering, Düzce University, Düzce, Turkey
4Industrial Engineering, Istanbul Medipol University, Istanbul, Turkey

Abstract

In recent years, researchers have highlighted how large volumes of data

can be transformed into information to determine customer behaviors,
and data mining applications have become a major trend. It has become
critical for organizations to use a tool for understanding the relationships
between data to protect their marketplace by increasing customer loyalty.
Thanks to data mining applications, data can be processed and
transformed into information, and in this way, target audiences can be
determined while developing marketing strategies. This chapter aims to
increase the market share with products specific to the customer
portfolio, introduce strategic marketing tools for retaining old customers,
introduce effective methods for acquiring new customers, and increase
the retail sales chart, based on purchasing habits of customers. The data
set was collected under pandemic conditions during the COVID-19
process and analyzed to support retail businesses in their online shopping
orientation. By examining the local customer base, it was assumed that

*
Corresponding Author’s Email: [email protected].

In: The Future of Data Mining

Editor: Cem Ufuk Baytar
ISBN: 979-8-88697-250-4
© 2022 Nova Science Publishers, Inc.
30 Elif Delice, Lütviye Özge Polatlı, İrem Düzdar Argun et al.

the customer group would display similar behaviors in online or tele-

ordering methods, customer identification and order estimation were
made to follow an effective sales policy. Segmentation was performed
with data mining applications, and the grouped data were separated
according to their similarities. The data set consisting of demographic
characteristics and various product information of the enterprise’s
customers were analyzed with Decision Tree and Random Forest, which
are data mining methods, the best performing algorithm in the data set
was selected by comparing the performance of the methods. As a result
of the findings, appropriate suggestions were given to the business to
determine the purchasing tendencies of the customers and to increase the
level of effectiveness in sales-marketing strategies. In this way, materials
were presented to assist the enterprise in developing strategies to increase
the number of sales by taking faster and more accurate action by avoiding
the time and expense that would be lost by the trial-error method.

Keywords: data mining, decision trees, random forest, k-means, x-means

Introduction

Along with the developing world, globalization, which has occurred with the
contribution of technology, has led to some changes on a sectoral basis and
has led to an increase in competition. The most important change that has
occurred is the evolution of customers’ purchasing tendencies and
expectations in a different direction over time. The differentiations, starting
from the customer, did not only stay with the customer, but also caused
different major effects in the continuation of the process with the domino
effect. The most critical point in the flow is to be able to manage the effects of
change by adopting an agile structure, considering the business vision and
mission. An agile management style, on the other hand, is directly proportional
to grasping the differences and keeping up with the change. Since the customer
is at an important point, efficient management of customer relations will have
a direct impact on the market share and will enable institutions to stay
competitive in the emerging competitive environment (Larson and Chang,
2016). At this point, data mining methods, which are frequently encountered
in applications in terms of their increasing importance and convenience in
recent years, attract the attention of researchers in sectoral as well as in
different fields and facilitate the management of processes.
Data mining is an analysis technique that is supported by algorithms to
reveal the information hidden in large amounts of data, plays a role in
Applications of Data Mining Algorithms … 31

predicting the future, and contains many different methods. Thus, by

highlighting useful and meaningful information, it sheds light for researchers
on related issues and adds efficiency to processes. At this point, service
providers need to keep up with the ongoing change and use it to their
advantage by extracting useful information from the big data formed to adopt
an agile structure. Understanding the changing customer expectations with
different data mining methods and taking steps according to the results brings
companies to the fore in their own fields of activity in differentiation that starts
from the customer and affects the whole process. While the methods under the
umbrella of data mining are basically grouped under two headings: methods
used for estimation and methods used for identification purposes, all the
methods have different features and functions. With these aspects, they can
answer more than one question at the same time, and they are becoming more
popular day by day (Gibert et al., 2018).
In this context, it is aimed to define the customer group in the data set by
using the clustering algorithms k-means and x-means. After defining the
customer group, it is aimed to examine the performance of different algorithms
on the data set with the support of decision tree algorithms. The most
important reason for choosing such a flow is to first subject the data to a
segmentation process and to give a reference point for classification in
decision tree algorithms. Thus, it was desired to determine the purchasing
tendencies of the customers by adding another column to the data set those
states which data belongs to which cluster, and by specifying how the
estimation will be made. As a result, a model that will enable the service
providers to take more robust and faster action will be put forward by
eliminating the time that companies will lose with trial-and-error methods.

Literature Review

Data mining has become increasingly popular in recent years, with

applications in practically every industry. The most important reason for this
is that all the methods under the umbrella of data mining enable to predict the
future with their different functions and make processes efficient. With the
spread of method applications, researchers turned their focus to this subject,
thus, the number and variety of studies in the literature have increased day by
day. Studies in which data mining application areas are explained have been
made in a wide range from specific applications and analyzes. In some
32 Elif Delice, Lütviye Özge Polatlı, İrem Düzdar Argun et al.

publications, the methods are examined alone, while in others, the

performances of the methods are compared.
For example, Pelleg and Moore (2000) made a comparison of two
different clustering algorithms by examining X-Means and K-Means methods
in their study. As a result, they revealed the methodological differences of the
methods discussed together with their reasons. On the other hand, Emel and
Taşkın (2005) focused on sales data with the CART decision tree technique in
their research, and as a result, they produced a study examining the estimation
success of the related method. Similar to Pelleg and Moore, Balabantaray,
Sarma and Jha (2015) compared the usage areas/differences of K-Means and
K-Medoids algorithms, which are the two most used methods among
descriptive data mining methods, using WEKA software. Also, Kale and
Yüksel (2020) conducted a study with different decision trees, but unlike Emel
and Taşkın, they compared the performance of seven different methods
instead of one method. In addition to all these studies, research conducted by
the authors focusing on different methods of data mining are presented in more
detail in the continuation of the chapter in order to show the different aspects
of the methods and to prove their efficiency.
Over the years, the subject has started to become popular and the number
of studies in the literature has increased. Thomas, Vinod and Raj (2014) aimed
to classify spam documents by developing a computationally efficient
classifier. Twelve feature selection techniques, namely TFDF, MI, PMI, NMI,
CDM, WMI, Chi-square, NGL, GSS, CPD, Fisher Score and LTC, were
applied and analyzed. As a result, it has been commented that RF with
Symmetric Uncert FST model is better for classification of spam documents.
Additionally, Guftar et al. (2015) aimed to present a new framework for
estimating the possible causes of syncope. They were analyzed a dataset from
Armed Forces Institute of Cardiology and the National Institute of Heart
Disease (AFIC & NIHD from Rawalpindi, Pakistan, using RapidMiner. In
conclusion, the proposed framework has been proven to improve prediction
accuracy through the new clustering approach for possible causes of syncope.
Data mining has continued to find its place in every field and has shown its
productive side to everyone.
There are some other studies that focused on different situations. Naik and
Samant (2016) classified 10 independent variables using liver patient data
from Indian liver patients. In this chapter, in which a total of 583 samples were
classified, Decision tree, Decision Stump, K-Nearest Neighbor and Naïve
Bayes algorithms were used. They used the programs WEKA, Rapid miner,
Tanagra, Orange and Knime. In the case of decision tree evaluation, the
Applications of Data Mining Algorithms … 33

Orange tool showed lower accuracy, while the Knime tool predicted better
accuracy compared to its predecessor. And also, Wu et al. (2017) examined
the purchasing behavior of more than 500,000 customers of the insurance
company from China Life Insurance Company for 3 years. Researchers used
SMOTE-based algorithms to analyze business data, in this way, they proposed
the Random Forest algorithm using the parallel computing capability and
memory caching mechanism optimized by Spark.
There are crises in the stock market that directly affect the economy.
Therefore, it is very important to identify early warning indicators. Elagamy,
Stanier and Sharp (2018) applied text mining and data mining applications to
identify these critical indicators. Random Forest, Rotation Forest, Bagging,
J48, Bayes Net, Decision table, Decision stump algorithms were used to
classify the critical indicators of the stock market. The best performing
algorithm was Random Forest algorithm. Random Forest, which was used in
the classification of the dataset, provided 98.34% accuracy and correctly
classified 535 of 544 articles. Likewise, Tan, Yan and Zhu (2019) aimed to
predict the long-term and short-term price trend by using the data of the
Chinese stock market between February 8, 2013 and August 8, 2017. In the
same context, Livari and Ghalam (2020) aimed to group customers for a food
manufacturing company using the variables of 2021 new purchases (R),
frequency of purchases (F) and monetary value of purchase (M). By
examining these variables, they used the K-Means algorithm and the Davies-
Bouldin Index algorithms to cluster the customer group of the seedling
production company. Customers were divided into three groups and marketing
strategies were suggested on the basis of each group by calculating the
customer lifetime value of each group.
In recent years, the number of studies has increased and data mining
methods have become an important tool in predicting the future.
Arminarahmah et al. (2021) aimed to view and map the spread of the COVID-
19 virus in Asia with a dataset containing information on the basis of total
cases, total deaths, total recovered and active cases from 2020, in 49 countries.
They completed the research using data from the site such as WHO, CDC,
NHC and worldometer. They used X-Means clustering method using
RapidMiner. It was interpreted that the X-Means algorithm could be used to
map the spread of COVID-19 in Asia, and it was divided into 4 regions
according to the results of the spread mapping analysis. In different
perspectives, there are some other studies conducted. Influencers promote
their products using social media accounts such as Twitter, Facebook, and
Instagram. Using the data collected from the marketing agency in Korea
34 Elif Delice, Lütviye Özge Polatlı, İrem Düzdar Argun et al.

includes purchase information such as customer information, purchase item

and payment amount from August 2018 to October 2020, Kim and Lee (2021)
aimed to predict customer churn in influencer trading. They used the computer
software program RapidMiner to predict lost customers. They achieved 90%
accuracy by applying the Decision Tree algorithm. Also, Utomo (2021) aimed
to analyze the data received from the Indonesian Ministry of Health on January
1, 2021 with using the clustering method. This chapter, which uses 2 attributes
as confirmed and death cases, also includes information such as confirmed
cases, treatment, recovery, and death cases. They compared the K-Means
method with K-Medoids to cluster the spread of the coronavirus in Indonesia,
as a result, it is seen that the best performing algorithm was the K-Means
algorithm.
In addition, other studies carried out in 2021 can be summarized as
following. Abdulkareem et al. (2021) aimed to observe the COVID-19 vaccine
progress in the world by using machine learning classification algorithms. It
has been discovered that the Decision Tree algorithm outperforms other
algorithms in terms of accuracy and time criteria. In order to create sales
clusters, Fithri and Wardhana (2021) aimed to analyze sales data with the
support of the K-Means Clustering algorithm. As a result, 3 sales clusters were
created, and it was suggested that sales cluster information could be an input
for an alternative solution, inventory management and marketing strategies.
Data clustering results from cluster1, cluster2 and cluster3 with 62%, 8%, and
30% percentage values, respectively. Evdokimova (2021) used segmentation
(ABC and XYZ analysis) and clustering methods (K-means, X-means,
Expectation-Maximization) in RapidMiner Studio to measure customer
engagement. They processed large amounts of data to analyze their customer
base. As a result, the best distribution of the customer base has been achieved
by the Expectation-Maximization method.
Studies carried out in different sectors using different methods and
software programs are presented in Table 1, and the algorithm with the best
performance is highlighted as a result of the study. In this context, the studies
summarized in the relevant table are shown on the basis of the author and year
of publication, problem considered in the study, compared algorithms,
algorithm that shows best performance, the accuracy of the algorithm and the
software program used.
Applications of Data Mining Algorithms … 35

Table 1. Studies carried out in different sectors using different

methods and software programs

best performance
Algorithm with

Program used
algorithms
Compared
Author(s)

Accuracy
Problem

(%)
Haghanikha Finding the highest Naïve Bayes, Decision 97 Orange,
meneh et al. performing tool Decision Tree, Tree Tanagra,
(2012) Support Vector KNIME,
Machine, K Nearest WEKA
Neighbor, One
Rule, Zero Rule
Alsultanny Forecasting the Naïve Bayes, Decision 100 -
(2013) needs of the labor Decision Tree, Tree
market Decision Rules
Thomas, Classify spam Naive Bayes, Random 98,73 WEKA
Vinod, and documents by Random Forest, Forest
Raj (2014) developing a Random Tree, J48,
computationally Adaboosting
efficient classifier
Naik and Classifying people Decision Tree, K-Nearest 99,70 WEKA,
Samant with and without Decision Stump, K- Neighbor RapidMiner,
(2016) liver Nearest Neighbor, Tanagra,
Naïve Bayes Orange,
Knime
Elagamy, Classification of Random Forest, Random 98,34 -
Stanier, and critical indicators of Rotation Forest, Forest
Sharp (2018) stock markets Bagging, J48,
Bayes Net,
Decision Table,
Decision Stump
Abdulkareem Observing COVID- Decision Tree, K Decision 99.9 WEKA
et al. (2021) 19 vaccine progress Nearest Neighbor, Tree
around the world Random Forest,
using machine Naive Bayes
learning
classification
algorithms

As a result, it has been observed that many studies in the literature are
supported by different methods under the umbrella of data mining in many
different fields. It has been observed that the studies have increased greatly in
2021, and that the methods under the umbrella of data mining are frequently
examined by researchers and used to facilitate processes. In this direction, it
has been determined that help is received from different software programs,
but RapidMiner and WEKA are the programs that come to the fore the most.
36 Elif Delice, Lütviye Özge Polatlı, İrem Düzdar Argun et al.

Methodology

Decision trees, one of the tree-based learning algorithms, are the most used
supervised learning algorithms. In general, they can be adapted to the solution
of all the problems (classification and regression) considered. Therefore, it is
very important for data analysts to learn and use these algorithms. The first
cells of the decision trees are called the root or root node. Each observation is
classified as “Yes” or “No” according to the root condition. Below the stem
cells are interval nodes or nodes. Each observation is classified with the help
of nodes. As the number of nodes increases, the complexity of the model also
increases. At the bottom of the decision tree is the leaf nodes or leave. The
leaves give us the result.
The Random Forest algorithm is a supervised classification algorithm.
There are two stages in the Random Forest algorithm. The first of these stages
is to create a Random Forest, and the other is to make predictions on the
Random Forest classifier created in the first stage. The K-means algorithm
provides an advantage in terms of being used in large data sets and providing
ease of application. In this method, the number of clusters should be defined
at the beginning, and it is not recommended to be used in categorical data sets.
One of the disadvantages of the K-means algorithm is that it cannot give
accurate results in data sets that differ in density and size. In addition, when
factors such as unusual data and noise occur, k-means remains weak compared
to other algorithms (Douzas, Bacao and Last, 2018). When using the K-means
method, a local optimum point is calculated, and the method ends at this point
(Ahmadyfard and Modares, 2008). Although the method is an iterative
method, it aims to divide the square of the distance between the elements into
the specified set using the minimization function. In each refresh, the data is
sent to a different cluster and the optimum result is tried to be found. At this
stage, a permutational operation is performed (Omran, Salman and
Engelbrecht, 2006).
The X-means algorithm answers the question of how to find the number
of k clusters by finding the optimum number. For this, it should be ensured
that the number of K clusters is defined in a reasonable range. The number k
in the specified range is scored by a model such as the Bayesian information
criterion. In summary, it is examined from the lower limit to the upper limit
of the determined range and works by adding a new center at the required
point. The centroid with the best score among the k numbers scored in this
process is considered optimum (Pelleg and Moore, 2000). The fact that the
Applications of Data Mining Algorithms … 37

determined clustering algorithms have different features and processing

methods, as stated above, is also included in the literature.
Cross Validation is used to evaluate the performance of predictive models
and to prevent underfitting and overfitting when it comes to fit. Berrar (2018)
performs performance measurement by separating the cross-validation dataset
into two separate groups and using one of them as a training set and the other
as a test set (Duda et al., 2005).
It also emphasized the importance of criteria while explaining the
classification concept and formulation. It is not possible to think that one of
the methods used in solving classification problems is better or worse than the
other. At this point, it is necessary to act according to the suitability of the data
set to the algorithm. There are established criteria to evaluate the performance
of classification algorithms. At the beginning of these criteria, it can be defined
as precision, Accuracy / Error Ratio, Specificity, Recall, F-measure, and ROC
area. Confusion matrix is used to calculate the criteria. Shown as Confusion
Matrix CMi,j, where i and j values are classes. CMi,j states that instances
tagged in class j actually belong to class i. In the following example there is
an example of a matrix with 2 different classes (Chandrashekar and Lee,
2019).
There are some terms and formulations used to interpret the outputs from
data mining methods. Based on this in the software developed in the chapter,
accuracy, sensitivity, and specificity values for classification problems are
calculated with the formula given in Equations 1, 2 and 3, respectively.

𝑇𝑃 + 𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (1)
𝑇𝑃 + 𝐹𝑁 + 𝐹𝑃 + 𝑇𝑁

𝑇𝑃
Precision = 𝑇𝑃 + 𝐹𝑃
(2)

𝑇𝑁
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = (3)
𝑇𝑁 + 𝐹𝑁

(𝑇𝑃 + 𝑇𝑁)
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 𝑅𝑎𝑡𝑒 = (𝑇𝑃 (4)
+ 𝐹𝑁 + 𝐹𝑃 + 𝑇𝑁)

(𝐹𝑁+𝐹𝑃)
𝐸𝑟𝑟𝑜𝑟 𝑅𝑎𝑡𝑒 = (𝐹𝑃+𝐹𝑁+𝐹𝑃+𝑇𝑁)
(5)

True Positives (TP): These are instances where the true value is 1 and the
predicted value is 1.
38 Elif Delice, Lütviye Özge Polatlı, İrem Düzdar Argun et al.

True Negatives (TN): These are instances where the true value is 0 and
our predicted value is 0.
False Positives (FP): These are instances where the true value is 0 but the
predicted value is 1.
False Negatives (FN): These are instances where the true value is 1 but
our predicted value is 0.
It is desired to measure the true negative performance with the specificity
criterion calculated in the third equation. It is the case that data that does not
belong to class C are not defined in class C.
Precision refers to the ratio of positive samples classified in the correct
category by the model to the total number of positive samples. In other words,
it is the rate at which the data estimated as C is actually C. The precision
measure measures the performance of the model in eliminating false positive
data. The nominal value is the ratio of correctly classified samples belonging
to the C class to the total C class. The rated measure measures the performance
of eliminating false negative values, unlike the precision criterion.

𝑇𝑃
Recall = 𝑇𝑃 + 𝐹𝑁
(6)

It is requested and expected that these two criteria give results close to 1.
The harmonic means of precision and recall measures is expressed as F
and calculated as follows.

2 𝑥 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 𝑥 𝑅𝑒𝑐𝑎𝑙𝑙
𝐹 = (7)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙

When it is desired to be calculated by giving weight according to the

precision criterion and the recall criterion, the 𝐹𝑏 value is calculated as
follows.

(1+𝑏 2 ) 𝑥 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 𝑥 𝑅𝑒𝑐𝑎𝑙𝑙

𝐹𝑏 = 𝑏 2 𝑥 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 +𝑅𝑒𝑐𝑎𝑙𝑙
(8)

After calculating these values, the interpretation is more successful than

the classifiers for which these values give larger results.
Applications of Data Mining Algorithms … 39

Findings and Results

Since 2019, the shopping habits of consumers have changed in accordance

with the living conditions brought by the pandemic worldwide. Logistics/
supply issues have gained importance with the concepts of online ordering and
e-commerce. In this process, companies have directed business development
studies, process shortening studies and activities to increase service quality in
order to respond to demands faster by taking the road to be covered in a shorter
time under normal conditions. The data set collected during the shopping of
the customers of a business working in the retail sector and used in this chapter
was created in August 2020 under pandemic conditions. In different time
periods, in the take-away customer base, in addition to quantitative data such
as age, gender, shopping unit, total amount, total weight, total shopping time,
how many people are involved in the shopping; the information on which
products they bought and the stability status of the customers from the moment
they entered the store were examined. The related data set contains data in
both string and integer formats. This data set, created with 213 customers and
16 different variables, allows working with almost all data mining methods.
Using this data set, the intended solution is to determine the purchasing
tendencies of customers with similar characteristics by looking at the
characteristics of take-away customers given in the data set, and to achieve
positive momentum that will affect the sales graph. Data mining studies are
carried out using software specially developed for this subject, as the data is
large-scale, as observed in the studies in the literature. SPSS, R, WEKA,
Python, Orange, RapidMiner are the main ones. RapidMiner is an open-source
program written in Java and can run on any operating system. In this chapter,
it was preferred to use RapidMiner software by looking the last studies
realized in this area because of its advantages.
The data set was tested with clustering and decision tree algorithms, in
this way, the success rates of the algorithms on the data set were determined.
First of all, the data set is run with various clustering algorithms, it is aimed to
divide the customers into the most meaningful and most suitable clusters.
Thus, various experiments were made, and it was seen that the distribution of
the data was optimally distributed when k = 10 according to the characteristic
of the data set. In Table 2, it was seen that the results of the two algorithms
were close to each other, and a detailed examination was made. In Table 2, it
has been determined which class the data clustered in the K-Means algorithm
is included in the X-Means algorithm. For example, it has been seen that the
data classified as cluster 2 in the K-Means algorithm are customers with id
40 Elif Delice, Lütviye Özge Polatlı, İrem Düzdar Argun et al.

numbers corresponding to elements 8 of cluster 9 and element 5 of cluster 4

according to X-Means. In addition, according to K-Means, all elements of
cluster 1 correspond to all elements of cluster 0 in X-Means. This proves that
the two algorithms classify on the same basis at certain points. Based on all
these, it can be interpreted that the two algorithms collect the data in similar
groups.

Table 2. K-Means X-Means cluster comparison and classification

equivalent of K-Means algorithm in X-Means algorithm

K means X Means K means X Means

Cluster 0: 39 items Cluster 0: 5 items Cluster 0 (39) Cluster 5 (39/45)
Cluster 1: 5 items Cluster 1: 15 items Cluster 1 (5) Cluster 0 (5/5)
Cluster 2: 13 items Cluster 2: 33 items Cluster 2 (13) Cluster9 (8/10), Cluster
4 (5/6)
Cluster 3: 17 items Cluster 3: 6 items Cluster 3 (17) Cluster 1(15/15),
Cluster 9 (2/9)
Cluster 4: 17 items Cluster 4: 6 items Cluster 4 (17) Cluster 8 (17/38)
Cluster 5: 7 items Cluster 5: 45 items Cluster 5 (7) Cluster 3 (6/6), Cluster
7 (53/53)
Cluster 6: 55 items Cluster 6: 2 items Cluster 6 (55) Cluster 2(2/33), Cluster
7 (53/53)
Cluster 7: 2 items Cluster 7: 53 items Cluster 7 (2) Cluster 7 (2/2)
Cluster 8: 27 items Cluster 8: 38 items Cluster 8 (27) Cluster 5(6/45), Cluster
8 (21/38)
Cluster 9: 31 items Cluster 9: 10 items Cluster 9 (31) Cluster 2 (31/33)
Total number of Total number of Total (213) Total (213)
items: 213 items: 213

Figure 1. K-Means and X-Means dataset distribution.

Applications of Data Mining Algorithms … 41

Figure 2. K-Means and X-Means parameters.

As can be seen in Table 2, as a result of the classification of the two

algorithms, 205 out of 233 data are distributed in the same classes in the K-
Means and X-Means algorithm. At this point, when a detailed examination
was made, it was observed that the data classified in the K-Means algorithm
was included in the X-Means algorithm, as in Table 2, and it was observed
that even though the names of the classes were changed, it was still in the same
class with 88% probability. In Figure 1, point graphs of these classifications
are given. When examined visually, the data distribution supports the results
in Table 2. Despite the 14 features in the data set, it has been determined that
the operators in the RapidMiner program give lower priority when the
clustering algorithms are applied, while some parameters stand out according
to their weights to the result. These parameters, which have a low effect on the
result, were neglected in the classification process, instead of all 14 features,
the chapter was carried out on the parameters seen in Figure 2. The methods
used in clustering analysis process data based on similar distances. Methods
such as Standardized Euclidean, Square Euclidean, Euclidean, Canberra,
Manhattan Mahalanobis, Minkowski can be used to determine these distances.
Since each method has a different technique, it is necessary to be careful in
the application of cluster analysis. As seen in Figure 1, the basis of the
clustering algorithm is that each element has similar characteristics with the
other elements in its own group and it has features that distinguish the groups
42 Elif Delice, Lütviye Özge Polatlı, İrem Düzdar Argun et al.

from each other. The K-Means algorithm provides an advantage in terms of

being used in large data sets and providing ease of application. In this method,
the number of clusters must be defined initially. When using the K-Means
method, a local optimum point is calculated, and the method ends at this point.
Although the method is an iterative method, it aims to divide the square of the
distance between the elements into the specified set using the minimization
function. In each refresh, the data is sent to a different cluster and the optimum
result is tried to be found. At this stage, a permutational operation is
performed.
The X-Means algorithm answers the question of how to find the number
of k clusters by finding th optimum number. For this, it should be ensured that
the number of k clusters is defined in a reasonable range. The number k in the
specified range is scored by a model such as the Bayesian information
criterion. In summary, this algorithm works by examining from the lower limit
to the upper limit of the determined range and adding a new center at the
required point. The centroid with the best score among the k numbers scored
in this process is considered optimum. The determined clustering algorithms
have different features and processing methods within themselves. The
resulting methodological differences as shown in Figure 1 directly affect the
results and cause the data to be divided into different classes. Figure 2 shows
the parameters used for the K-Means and X-Means algorithms.
Clusters in the results of the K-Means algorithm, one of the clustering
methods, were included in the data set and studied with the help of Decision
Tree and Random Forest, which are Decision Tree methods, and supported by
Cross Validation due to the small amount of data. Cross Validation is used to
evaluate the performance of predictive models and to prevent underfitting and
overfitting when it comes to fit (Berrar, 2018). Cross-validation divides the
data set into two separate groups and uses one of them as a training set and the
other as a test set and performs performance measurement (Duda et al., 2005).
There are established criteria to evaluate the performance of classification
algorithms. At the beginning of these criteria are precision, accuracy / error
rate, specificity, recall, F-measure, and ROC area. Confusion matrix is used to
calculate the criteria. Confusion Matrix is denoted as CMi,j where i and j
values are classes. CMi,j indicates that the instances tagged in class j actually
belong to class i. In the following example there is an example of a matrix
with 2 different classes.
As can be seen in Table 3, the matrix is divided into 4 areas, and the values
of these areas are determined according to whether the data is classified
correctly or incorrectly. Field A is named as true positive, and it means that
Applications of Data Mining Algorithms … 43

the data of the C1 class in the training set is defined as C1 again. (B field false
negative) Represents data that is not defined as C1 in the test set, although it is
in the C1 class in the training set. The C false positive area is the area not
defined as C1 but classified as C1 when estimating. The D negative field is the
field that is not in the C1 class and is not predicted as C1. By using these values,
calculations related to accuracy and error rate were made.
In Table 4 and Figure 3, the estimation performances of the model applied
on the test data are given. In Figure 3, correctly predicted data are marked in
orange and incorrect ones in gray. Class precision and class recall values in
row and column subtotals are given in Table 3. From these values, with class
precision, the correct classification percentage of the Decision Tree algorithm
is calculated as 81.13% and the average of Recall values is 71.70%. In the
Random Forest algorithm, the percentage of correct classification is 85.71%
and the average of Recall values is 88.33%. As can be understood from here,
the Random Forest model performed better in estimation than the Decision
Tree model.

Table 3. The structure of the confusion matrix

Predicted Values
Positive (C1) Negative (-C1)
Actual
Positive (C1) A: True Positive (TP) B: False Negative (FN)
Values
Negative (-C1) C: False Positive (FP) D: True Negative (TN)

Table 4. Comparison of the success rate of algorithms

Decision Tree Class precision average 81.13%
Class Recall average 71.70%
Random Forest Class precision average 85.71%
Class Recall average 88.33%

When choosing the CART algorithm to distribute the model in the

Decision Tree model, all the predicted values are the same, which means that
this model has poor performance. Instead of creating a single tree, the Random
Forest model creates several trees that branch out all the nodes based on the
best of the randomly picked features at each node.
Datasets are produced as a result of the displacement of the data in the
dataset. No pruning is done, and random trees are developed. In decision trees,
a node represents the entire instance because it is the test performed for a
feature. The branch represents the result of the test performed, and the leaves
represent the classes as undivided nodes. The root node is the top node and it
is a model that works from top to bottom.
Figure 3. Decision Tree and Random Forest output.
Figure 4. Decision Tree obtained by Decision Tree algorithm and Random Forest.
46 Elif Delice, Lütviye Özge Polatlı, İrem Düzdar Argun et al.

With the Decision Tree, we see that the first node is separated according
to the total weight, while in the Random Forest method, we see that the first
node is the purchase time. Images of decision trees obtained by Decision Tree
and Random Forest algorithms are given in Figure 4. Although the Random
Forest model, from the way it is expressed with colors under the clusters in
the images, performs better in estimating, it is seen that the criterion of the
homogeneity of the subset in the decision trees stands out more in the Decision
Tree. As a result, with all these findings, it was seen that the Random Forest
algorithm performed more successfully than the Decision Tree algorithm in
the analysis on the data set processed and specified in the RapidMiner
program.

Conclusion

Since 2019, consumers’ shopping habits have changed due to limited living
conditions around the world. The data set used in this chapter consists of data
collected under pandemic conditions in August 2020. The aim was to form a
basis for the online shopping activities of retail businesses under pandemic
conditions. By examining the local customer base and assuming that the same
customer group will exhibit similar behaviors in online or tele-ordering
methods, it was desired to work on customer identification and order
estimation in order to follow an effective sales policy.
At the beginning of the chapter, first of all, clustering algorithms were
used for customer segmentation, and it was aimed to group the customers with
similar characteristics. Thus, it is aimed to reach the result in a shorter time by
including the customers with certain characteristics, rather than each customer
being a value on its own.
In order to make predictions on the customers divided into clusters,
decision tree algorithms were tried, and it was investigated which algorithm
would perform higher in the relevant data set. The aim here was to develop
effective sales strategies by looking at the branches of the algorithm with the
highest performance.
The effect of the existing differences in the functioning of the selected
methods on the result was observed, and the reasons for which method was
appropriate were investigated. At this point, the output of the chapter is an
exemplary chapter, in addition to the existing information in the literature, in
terms of reference to studies that will be compared in performance on similar
data sets in which integer and string values are handled in a mixed manner.
Applications of Data Mining Algorithms … 47

As stated in the findings section, the randomness of the Random Forest

model and the creation of multiple trees in the chapter created a positive value,
resulting in a higher performance than the Decision Tree method, which works
on estimating a single tree and the same data.
In order to make additional predictions to the algorithms studied for the
future, the effectiveness of models such as Chadi, ID3, logistic regression,
gradient boosted tree can be tested, and comparisons can be made. It will help
to determine the most suitable estimation model for the data set by
investigating the causality in the performance differences. At the same time,
causality can be investigated by observing the differences between the coding-
based programs and the programs working with the help of operators by means
of WEKA, SPSS, R language, Python, Orange, which are used in data mining
studies.

References
Abdulkareem, N. M., Abdulazeez, A. M., Zeebaree, D. Q., and Hasan, D. A. (2021).
COVID-19 World Vaccination Progress Using Machine Learning Classification
Algorithms. Qubahan Academic Journal, 1 (2), 100–105.
Ahmadyfard, A., and Modares, H. (2008). Combining PSO and K-Means to Enhance Data
Clustering. 2008 International Symposium on Telecommunications, IST 2008, 688–
691.
Alsultanny, Y. A. (2013). Labor Market Forecasting by Using Data Mining. Procedia
Computer Science, 18, 1700–1709.
Arminarahmah, N., Daengs, A., Bhawika, G. W., Dewi, M. P., and Wanto, A. (2021).
Mapping the Spread of Covid-19 in Asia Using Data Mining X-Means Algorithms.
IOP Conference Series: Materials Science and Engineering, 1071 (1), 012018.
Balabantaray, R. C., Sarma, C., and Jha, M. (2015). Document Clustering Using K-Means
and K-Medoids. http://arxiv.org/abs/1502.07938.
Berrar, D. (2018). Cross-Validation. Encyclopedia of Bioinformatics and Computational
Biology: ABC of Bioinformatics, 1–3, 542–545.
Chandrashekar, M., and Lee, Y. (2019). MCDD: Multi-Class Distribution Model for Large
Scale Classification. Proceedings - 2018 IEEE International Conference on Big Data,
Big Data, 2018, 4906–4914.
Douzas, G., Bacao F., and Last, F. (2018). Improving Imbalanced Learning through a
Heuristic Oversampling Method Based on K-Means and SMOTE. Information
Sciences, 465, 1–20.
Duda, S., Miller, A. R., Statnikov, A., and Johnson, K. (2005). Extracting Drug-Drug
Interaction Articles from MEDLINE to Improve the Content of Drug Databases. AMIA
Annual Symposium Proceedings/AMIA Symposium. AMIA Symposium, 216–220.
48 Elif Delice, Lütviye Özge Polatlı, İrem Düzdar Argun et al.

Elagamy, M. N., Stanier, C., and Sharp, B. (2018). Stock Market Random Forest-Text
Mining System Mining Critical Indicators of Stock Market Movements. 2nd
International Conference on Natural Language and Speech Processing, ICNLSP
2018, 1–8.
Emel, G. G., and Taşkın, Ç. (2005). Veri Madenciliğinde Karar Ağaçları ve Bir Satış
Analizi Uygulaması [Decision Trees in Data Mining and a Sales Analysis
Application], 6 (2), 221–239.
Evdokimova, S. A. (2021). Segmentation of Store Customers to Increase Sales Using ABC-
XYZ-Analysis and Clustering Methods. Journal of Physics, Conference Series, 2032
(1).
Fithri, F. A., and Wardhana, S. (2021). Cluster Analysis of Sales Transaction Data Using
K-Means Clustering At Toko Usaha Mandiri. Jurnal PILAR Nusa Mandiri, 17 (2),
113–118.
Garcia, C. D. S., Meincheim, A., Junior, E. R. F., Dallagassa, M. R., Sato, D. M. V.,
Carvalho, D. R., Santos, E. A. P., and Scalabrin, E. E. (2019). Process Mining
Techniques and Applications – A Systematic Mapping Study. Expert Systems with
Applications, 133, 260–295.
Gibert, K., Izquierdo, J., Sànchez-Marrè, M., Hamilton, S. H., Rodríguez-Roda, I., and
Holmes, G. (2018). Which Method to Use? An Assessment of Data Mining Methods
in Environmental Data Science. Environmental Modelling and Software, 110, 3–27.
Guftar, M., Ali, S. H., Raja, A. A., and Qamar, U., (2015). A Novel Framework for
Classification of Syncope Disease Using K-Means Clustering Algorithm. IntelliSys
2015 - Proceedings of 2015 SAI Intelligent Systems Conference, 127–132.
Haghanikhameneh, F., Panahy, P. H. S., Khanahmadliravi, N., and Mousavi, S. A. (2012).
A Comparison Study between Data Mining Algorithms over Classification
Techniques in Squid Dataset. International Journal of Artificial Intelligence, 9 (12 A),
59–66.
Kale, B., and Yüksel, G., (2020). Veri Madenciliği Sınıflandırma Algoritmaları Ile E-Posta
Önemliliğinin Belirlenmesi. Ç.Ü Fen ve Mühendislik Bilimleri Dergisi
[Determination of E-Mail Importance with Data Mining Classification Algorithms.
CU Journal of Science and Engineering Sciences], 39 (9).
Kim, S., and Lee, H. (2021). Customer Churn Prediction in Influencer Commerce: An
Application of Decision Trees. Procedia Computer Science, 199, 1332–1339.
Larson, D., and Chang, V. (2016). A Review and Future Direction of Agile, Business
Intelligence, Analytics and Data Science. International Journal of Information
Management, 36 (5), 700–710.
Livari, R. T. and Ghalam, N. Z. (2020). Customers Grouping Using Data Mining
Techniques in the Food Distribution Industry (A Case Study). SRPH Journal of
Applied Management and Agile Organisation.
Naik, A., and Samant, L. (2016). Correlation Review of Classification Algorithm Using
Data Mining Tool: WEKA, Rapidminer, Tanagra, Orange and Knime. Procedia
Computer Science, 85, 662–668.
Omran, M. G. H., Salman, A., and Engelbrecht, A. P. (2006). Dynamic Clustering Using
Particle Swarm Optimization with Application in Image Segmentation. Pattern
Analysis and Applications, 8 (4), 332–344.
Applications of Data Mining Algorithms … 49

Pelleg, D., and Moore, A. (2000). X-Means: Extending K-Means with Efficient Estimation
of the Number of Clusters. ICML '00: Proceedings of the Seventeenth International
Conference on Machine Learning.
Tan, Y., Chen, H., Zhang, J., Tang, R., and Liu, P. (2022). Early Risk Prediction of Diabetes
Based on GA-Stacking. Applied Sciences (Switzerland), 12 (2).
Tan, Z., Yan, Z., and Zhu, G. (2019). Stock Selection with Random Forest: An Exploitation
of Excess Return in the Chinese Stock Market. Heliyon, 5 (8), e02310.
Thomas, J. P. V., and Raj, N. S. (2014). Towards Spam Mail Detection Using Robust
Feature Evaluated with Feature Selection Techniques. International Journal of
Engineering and Technology, 6 (5), 2144–2158.
Utomo, W. (2021). The Comparison of K-Means and k-Medoids Algorithms for Clustering
the Spread of the Covid-19 Outbreak in Indonesia. ILKOM Jurnal Ilmiah, 13 (1), 31–
35.
Wei, C. P., and Chiu, I. T. (2002). Turning Telecommunications Call Details to Churn
Prediction: A Data Mining Approach. Expert Systems with Applications, 23 (2), 103–
112.
Wu, Z., Lin, W., Zhang, Z., Wen, A., and Lin, L. (2017). An Ensemble Random Forest
Algorithm for Insurance Big Data Analysis. Proceedings - 2017 IEEE International
Conference on Computational Science and Engineering and IEEE/IFIP International
Conference on Embedded and Ubiquitous Computing, CSE and EUC 2017, 1, 531–
36.
Chapter 4

Analysis of Customer Churn in the Banking

Industry Using Data Mining

Özge Doğuç*
Management Information Systems, Medipol University,
Istanbul, Turkey

Abstract

Today, banks have a very important place in the great economic

environments of countries. As in every sector, there are many
competitors and a great competitive environment in the banking field.
Especially individual customers prefer digital channels to make their
banking transactions faster and easier. Banks need to take fast and
industry-leading steps to meet these expectations of their customers.
They need to differentiate themselves from the competition with
innovative features by giving importance to digital. The main goals of
the banks in the competitive environment are gaining new customers,
increasing customer loyalty, reducing customer churn rates, and
providing superior customer satisfaction. In this study, customer data
belonging to a bank were analyzed with data analysis algorithms.
Customer churn analysis was performed using different machine
algorithms. The model was created on the Knime platform. This study
performs a customer loss analysis using data mining algorithms. The aim
is to reveal the reasons for losing customers, the elements of customer
loyalty and to help develop customer relations activities accordingly.

Keywords: customer relations, customer loyalty, churn customers, data

analysis


Corresponding Author’s Email: [email protected].

In: The Future of Data Mining

Editor: Cem Ufuk Baytar
ISBN: 979-8-88697-250-4
© 2022 Nova Science Publishers, Inc.
52 Özge Doğuç

Introduction

Companies attach importance to customer loyalty in an increasingly

competitive environment. Acquiring new customers is more costly than
increasing the loyalty of an existing customer and keeping them from leaving.
Systems that predict customers who are likely to abandon will contribute
to marketing and customer loyalty efforts. By focusing on this customer group,
companies can solve the customer loss problem. In this study, it is aimed to
analyze customer churn by using three data mining algorithms.
Knime and Python are used in this study. Knime is an open source and
multiplatform software data analysis, reporting and integrating platform. It is
used in pharmaceutical research and analysis in the fields of CRM, business
intelligence and finance. In this study, it was used in the stages of
understanding the data and creating a model. Python language is a widely used
software language in the field of data mining. NumPy, SciPy, MatPlotLib,
SciKit-Learn, Pandas libraries are data mining libraries used for data
visualization, processing, regression and many more analysis. In this study,
the Python language was used in the stages of understanding, interpreting, and
visualizing the data set.
In this study, customer loss estimation was made by using Decision Tree
and Random Forest algorithms, which are classification techniques of
supervised learning algorithms, and Artificial Neural Network. The accuracy
of the models was found, and the models were compared. The accuracy values
of all three models were close to each other. The Random Forest algorithm
caught the highest value with 0.864. The current data set has been obtained
from a publicly available source on the internet. The value of the target (class)
variable in the data set we are trying to predict, the value of the customer
leaving is 1, and the equivalent of not leaving is 0. There was a number
imbalance between the two values in the data set. Therefore, data increase has
been achieved by using the SMOTE node. Thus, an increase in the accuracy
value was observed. For the algorithms to have the best success values,
experiments were made by giving different values to the hyperparameters we
gave. The data mining methodology, CRIPS-DM (Cross-Industry Standard
Process for Data Mining), has been followed. Understanding the problem and
data; cleaning the data, making it usable; The stages of modeling and model
evaluation and comparison of success values have been completed. A roadmap
was drawn for the final implementation phase.
Analysis of Customer Churn in the Banking Industry Using Data Mining 53

Digitalization and Online Banking

Definitions of digitization vary. Sassen (Sassen, 2000) stated that in terms of

economy, digitalization expands the borders of the country, thereby increasing
convenience and globalization. He stated that with the digitalization of the
economy and globalization, later on, concepts such as sovereignty, culture and
capital were affected. Today, digitalization has experienced changes for both
companies and societies and continues to be up-to-date. New technologies
have been adopted within the company, causing significant changes. The study
of Parviainen et al. (Parviainen et al., 2017) has set a right starting point and
presented a model that aims to help companies to incorporate digital
transformation in a systematic way. The stages of the company's
implementation of the roadmap in order to reach the targets, after the
digitalization positioning, determination of the targets, analysis of the current
digital situation, were conveyed to the companies. Although this process
differs for each company, the work of Parviainen et al. has provided guidance
on the steps companies should take in their digitalization journey (Parviainen
et al., 2017).
Digitalization has brought different approaches in the services provided
to customers by each sector. The relationship established with the customer in
banking was renewed. The bank's digital information systems operation had
to reach the customer through the transfer of branch personnel. With the online
banking interfaces, the banking transaction limits of the customers are self-
determined and spread to all possible places. The products of the banks have
turned into providing information technology services directly to the
customers by leaving the service they provide only in their branches. Online
banking has made it possible for the customer to access account and
transaction information, create orders, requests, and perform many services
directly at any time, such as money transfer according to needs (Gupta and
Kamilla, 2014).
Apart from cash withdrawal, internet banking allows customers to easily
perform all kinds of transactions. As people's lives get more intense, their
demands for products that are easier and more widely used have increased.
Using information technologies has gained importance to increase customer
loyalty and product usage. A customer's perceived trust in online banking
services is one of the factors affecting the customer's loyalty to the bank and
the degree of adoption. The level of trust the bank establishes with the
customer through the technological interface affects the customer's future
commitment to online banking (Mukherjee and Nath, 2003).
54 Özge Doğuç

Customer Relations Management and Data Analysis

Customer relationship management has become an area that benefits from and
is fed by data mining analysis due to the increasing technology, competitive
environment, and information age. Customer relationship management aims
to increase customer loyalty, to be in constant communication with the
customer and to provide the right service at the right time. These positive
relations with the customer are important in terms of reducing customer loss
rates and acquiring new customers in the company's competitive environment.
Using information systems and algorithms to make sound decisions about the
future will bring significant advances in all areas of the company (Payne,
2005).
As a result of today's technological developments, increasing data
collection and storage possibilities, producing meaningful information from
these data has become one of the issues that companies attach importance to.
The usage methods of this data may vary in each sector. The problem to be
solved, the data used, the algorithms used may be different for each company
and sector. The study by Doğan et al. (Doğan, Erol and Buldu, 2014) includes
the use of data analysis algorithms in terms of customer relationship
management in the insurance industry. This study belongs to the various
insurance policy sales made by one of the most important insurance companies
in Turkey between 2010 and 2012. The data consists of approximately 12
thousand lines and 9 columns containing various insurance information such
as motor insurance, fire, earthquake. To ensure personal information security,
the customer ID and number are hidden. Apriori algorithm was used. The
letters "T and F" are used to represent whether there is a fuse or not. After the
data pre-preparation, the analysis phase was started. The association was
analyzed, and significant results were revealed such as “47% of those who
purchased compulsory earthquake insurance also bought fire insurance.” In
the light of this information, it will be possible to organize campaigns that will
attract the attention of the customer in future campaigns. In the insurance
sector, it is possible to obtain meaningful information through data analysis
by using customer data and to use this information in customer relationship
management and marketing activities. As a result of this study, it has been
shown that by looking at the preferred policy types, strategies such as which
customer will be directed to which campaign in the future, and thus efficiency
can be increased.
Hsieh's study (Hsieh, 2004) includes analyzing bank customer data with
data mining and drawing conclusions. The databases are large as banks have
Analysis of Customer Churn in the Banking Industry Using Data Mining 55

daily and monthly transaction surpluses, registered accounts and a wide range
of customers. Data analysis studies are also challenging. This study was also
conducted on credit card customers and a behavioral scoring model was
created by analyzing the historical data of existing customers. Three main
profitable customer groups were separated by using variables such as
repayment behavior, frequency of use, duration of use and demographic
information. Apriori association algorithm was used. This study shows that
identifying customers with the behavioral scoring model provides useful
information and facilitates marketing strategy development. As a result, it has
been shown that credit and behavior scoring models are useful and this study
will be useful in making a more accurate decision on whether to give credit to
newly made loan applicants. It helped marketers determine efficient strategies
with customer profiles and conduct more personalized marketing customer
relations.
Gürsoy's (Gürsoy, 2010) study titled “Customer Dropout Analysis in the
Telecommunication Sector” studied churn customer loss with data analysis
algorithms in the telecommunication sector. If a customer cancels their
membership agreement with a company and becomes a customer of another
competitor, that customer is called a lost customer or a Churn customer. Loss
of customers is closely related to customer loyalty. Price advantages are not
enough to keep the customer. Adding new value-added services to products is
an important point to ensure customer loyalty. The main purpose of customer
churn analysis is to find a customer that is likely to be lost and to take
marketing or other strategic steps to avoid losing those customers. The lost
customers to be considered in the analysis differ according to the sectors or
data. For example, a credit card customer can easily start using another bank's
credit card without canceling the current bank's credit card. In this case, one
can look at the rate of reduction in expenses to understand the loss of the
customer. Loss of customers can be a big problem in competitive banks,
insurance, and telecommunications companies. The cost of acquiring new
customers for companies is increasing day by day. Instead of organizing
campaigns to gain new customers, companies care about higher customer
loyalty. Having a good model that handles large amounts of data will help
companies to plan competitive advantage and more accurate strategies. In the
study of Şimşek and Oman (Şimşek, 2018), customer loss estimation of a
telecommunication company in Turkey was made with SPSS. Identifying
which customers are likely to lose, determining which customers should
retain, and developing strategies to retain profitable customers are matters of
customer retention. Firstly, it is very important to determine the churn ratio.
56 Özge Doğuç

By using this ratio, companies can make predictions about the future behavior
of new customers and develop appropriate strategies beforehand.
Logistic regression and decision tree algorithms were used in this study.
The target variable in the data set, customer leaving and not leaving, is given
as 0, 1. After preprocessing the data, the split ratios of 0 and 1 were balanced
as 49% to 50%. The next step is to eliminate variables that have no effect on
the target variable. By examining the correlation between the variables, the
determined variables are eliminated. As a result of the decision tree model,
meaningful information such as “If the average of local and long-distance calls
is higher than 218, 42% of the subscribers will go to churn” was obtained. The
prediction accuracy rate of the logistic regression model that they did not
churn was 74%, and the accuracy rate of prediction that they did not churn
was 66%. As a result, the information obtained from this study can be used in
marketing activities. When campaigns are organized for products and services,
it can prioritize subscribers who are more likely to churn, thus reducing
customer churn (Gürsoy, 2010).
The study by Savaşçı and Tatlıdil (Savaşçı and Tatlıdil, 2006) includes
the evaluation of the practices of banks in customer relations in terms of
customer loyalty. Banks in the financial services sector aim to increase
customer loyalty, maintain the number of customers, and ensure satisfaction
by giving importance to customer relations. A customer data warehouse needs
to be created. In this study, customer impression is provided through credit
cards. Credit card strategies used for customer loyalty were evaluated. A
survey method was determined for 500 card users in İzmir district. The reasons
for using the most preferred credit cards by the consumers participating in the
survey were examined. When asked whether they use the installment facility
with their credit cards, 91% stated that they pay their purchases in installments.
The advantage with the highest level of satisfaction of the card(s) they use the
most is that it is wide and easy to use at a rate of 84%. Banks attach great
importance to customer relationship management to retain their customers and
increase customer loyalty and satisfaction in today's technology and
competitive environment. For this, it is important to keep customer data, to
obtain meaningful results by using data mining methods, and to use
communication technologies effectively and efficiently.
In the study of Gülpınar (Gülpinar, 2013), estimating customer loss in the
Turkish telecommunications sector with Artificial Neural Network analysis
and considering incoming outgoing calls, because of Social Network analysis,
it shows appropriate marketing activities for effective customers. A 14-
question survey was conducted with 100 different GSM users. For the Social
Analysis of Customer Churn in the Banking Industry Using Data Mining 57

Network analysis, the phone numbers of the 5 people they talked to last time
were asked and a survey was conducted with them. Thus, a wide
communication network was created. A high iteration model was created, and
the margin of error was reduced below 0.05. This study has added a different
approach to customer loss analysis by examining the communication network
structures of customers using social network analysis. What influences
customer value is not only their own behavior, but also the customer's network.
In the telecommunication sector, besides his personal characteristics, the
people he is connected to have affected the loss of customers. In other fields
such as banking, this approach to customer analysis will support the
emergence of better forecasting methods.
In a competitive telecommunications market, the company's goal of
minimizing customer loss by increasing customer loyalty is as important as
gaining customers. As studies have shown, spending on acquiring customers
is 20 times more costly than retaining existing customers. Customer
relationship management (CRM) establishes and manages strong, long-term
customer relationships with high loyalty. For this purpose, customer loss
analysis tools are developed and benefited from. (Vafeiadis et al., 2015)

Customer Loyalty and Data Analysis

It is important for companies to be able to estimate the amount of revenue they

can generate from their active customers. Therefore, companies need models
that will determine whether their customers are loyal to the company or not
and determine the number of customers who will leave the company and turn
to their competitors. The cost of acquiring a new customer is often higher than
the cost of retaining a customer. These models are especially important for the
modern telecommunication operator. In the study of Wijaya and Girsang
(Wijaya and Girsang, 2015), C4.5, Naive Bayes and Nearest Neighbor
Algorithms and data mining estimation of the national multimedia company
in Indonesia were studied. As a result, modeling customer loyalty degrees with
mathematical methods is beneficial for companies. These models will help the
company better forecast its revenues. With the development of database and
computing technology, data collection and storage processes have become
easier and faster. Thus, data mining processes increase and allow us to obtain
models with high success rates. The C4.5 algorithm was the algorithm with
the highest accuracy.
58 Özge Doğuç

The study of Ivanchenko et al. (Ivanchenko et al., 2019) on developing a

marketing relationship based on data mining in the banking sector will help
the sector in catching the advantages and opportunities in this field. This study
shows data mining methods and successful examples of Russian banks in this
field. It gives advice on the use of data mining in determining digital marketing
banking strategies. A Russian bank uses big data analytics to manage risks,
fight fraud, segment customers and evaluate customers' credit ratings, manage
personnel, segment customer churn, create financial statements, and analyze
posts on social networks and forums. The positive effects of providing
personalized offers based on outputs from data analysis for customers through
the Call Center have emerged. Contributions of using data analysis
technologies; individual customer satisfaction, service planning according to
customer needs, increase in customer loyalty and attracting the right target
customer with low costs.
The study by Maryani et al. (Maryani and Riana, 2017) aims to perform
clustering and customer profiling for mid-industry companies using the
Recency Frequency and Monetary (RFM) model for customer relationship
management. Customer segmentation, classification and determination of
customer loyalty levels were carried out with K-means and decision tree
algorithms.

Tools and Methodology

In this study, Decision Tree, Random Forest, and Artificial Neural Networks
data mining algorithms were applied on a data set containing bank customer
data. Data mining methodology, which stands for CRIPS-DM, is Cross-
Industry Standard Process for Data Mining. This is the most widely used
model in data mining and data science and has 6 stages.

• Business Understanding: It is the stage where the goals and needs of

the business are understood.
• Data Understanding: It is the stage of understanding the data. It starts
with data collection. It is the stage of understanding the data and
forming the first impression.
• Data Preparation: It is the stage of organizing the data set that we have
scattered in accordance with the model we will establish. Operations
Analysis of Customer Churn in the Banking Industry Using Data Mining 59

such as selecting data variables, cleaning data, placing missing data,

data type conversion can be done.
• Modeling: It is the stage of choosing various data mining algorithms
that we will apply to the data we have prepared, determining the
parameters, and applying the model.
• Evaluation: It is the stage of finding the evaluation scores of the
model such as accuracy and sensitivity by comparing the target
variable assumptions of the established model, which we call class,
with the test data. As a result of these trials, the model is reviewed,
and improvements are made if necessary.
• Implementation: It is the final stage. The model is presented to the
necessary users and brought to life in a way that it will be used in
accordance with its purpose (Alaybeg, 2019).

In this study, the data set consisting of 10,000 customers obtained from
open sources for customer churn analysis includes 14 variables. One of them
is the exit (Exited) target (class) variable that is tried to be estimated. These
are the variables;

• Registration id (RowNumber), Customer number (CustomerId),

Customer surname (Surname): These are the variables that have no
effect on the customer's departure from the bank. They did not
participate in modelling.
• Credit score (CreditScore): A customer with a high credit score is less
likely to leave the bank, which can have an impact on customer churn.
• Geography: customer location can have an impact on customer churn.
• Gender: It may be necessary to examine the effects on customer
churn.
• Age (Age): It influences the loss of customers. Older and younger
customers may have different behaviors.
• Usage period (Tenure): Refers to the number of years the customer
has been a customer of the bank.
• Balance: It is the customer's account balance amount, it influences
customer loss.
• Number of products (NumOfProducts): It is the number of products
purchased by the customer through the bank.
• Credit card ownership (HasCrCard): Indicates whether the customer
has a credit card. (0 as 1)
60 Özge Doğuç

• Activity (IsActiveMember): Indicates the customer's activity

according to the number of transactions. (in 0.1)
• Estimated Salary: People with low salaries may be more likely to
leave the bank than people with high salaries. It is a continuous
numeric variable.
• Lost (Exited): The bank customer's exit status is the target (class)
variable. (0: didn't leave, 1: left)

Understanding the Data

After understanding the purpose of the first stage of the CRIPS-DM

methodology, the stage of understanding the data was started. At this stage,
the Python programming language was used. The dataset was read, and tables
were extracted using the Seaborn and Matplotlib.pyplot libraries for statistical
graphs. Pandas, numpy library was used and data preliminary information was
obtained with appropriate codes. This phase has been completed by using
some nodes in Knime.

Table 1. Variable properties of age and credit score

Age (Age) Credit Score

Number (Count) 1000.000000 1000.000000
Average (Mean) 38,921800 650.528800
Standard deviation 10.487806 96.653299
(std)
Min value 18,000,000 350,000,000
First 25% average 32,000,000 584,000,000
First 50% average 37,000,000 652,000,000
First 75% average 44,000,000 718,000,000
Max value 92,000,000 850,000,000
Graphic

explanation It is seen that the extreme data is It is seen that the extreme data is
after the age of 60. before 400.
Analysis of Customer Churn in the Banking Industry Using Data Mining 61

Table 2. Variable properties of number of products and estimated salary

Number of Products Estimated Salary

(Num of Products)
Number (Count) 1000.000000 1000.000000
Average (Mean) 1.530200 100090.239881
Standard 0.581654 57510.492818
deviation (std)
Min value 1.000000 11.580000
Average of the 1.000000 51002.110000
first 25%
Average of the 1.000000 100193.915000
first 50%
Average of the 2,000,000 149388.247500
first 75%
Max value 4,000,000 199992.480000
Graphic

explanation There is a total of 60 people with 4 It is seen that the estimated salary
product ownership, seen as extreme distribution is balanced.
data. 1 and 2 have agglomeration.

The characteristics of the variables of Age, Credit Score, Number of

Products and Estimated Salary in the data set are given in Table 1 and Table
2.
The correlation relationship between the variables was examined. If there
were variables with a correlation ratio greater than 0.5, we would have taken
it into account. The highest values are between 0.3 on the negative side and
Balance and NumOfProducts. The correlation graph is given in Figure 1.
Correlation coefficients were examined with the corr() function whether
continuous numeric variables have a bilateral relationship on the customer
churn variable. Significant values close to 1.0 were not found. A few values
are shown in Table 3.
With the Interactive Histogram module in Knime, scatter charts of the
variables in terms of customer loss were drawn in Figure 2. Customers shown
in green have not left the bank (variable Exited = 0). Customers shown in red
are customers who have left the bank (variable Exited = 1).
62 Özge Doğuç

Figure 1. Correlation plot of variables.

Table 3. Variable correlation coefficients

Variables Correlation coefficient

Customer's account balance amount (Balance) 0.11853276877163386
Usage time (Tenure) -0.014000612253444594
Credit score (CreditScore) -0.02709353975465778

The distribution of the age variable in terms of customer loss is shown

with the histogram graph in Figure 3. We see that the number of customer
losses in the 42-50 age range is higher than other age ranges. Then comes the
34-42 age range, the 50-58 age range and the 26-34 age range. In the 50-58
age range, we see a higher rate of separation compared to other ranges.
Analysis of Customer Churn in the Banking Industry Using Data Mining 63

Figure 2. Loss of customers (0 = customer did not abandon,

1 = customer abandoned).

Figure 3. Customer loss distribution by age variable.

It is seen that the loss of customers is approximately evenly distributed in

the number of years the bank has been a customer in Figure 4. When we look
at the ratio, we see that the loss rate is lower for new customers between 0-2
years.
In Figure 5, it is shown that the number of customers with a balance of 0-
41900,001 is higher and the loss rate is low compared to other ranges.
64 Özge Doğuç

Figure 4. Distribution of number of years the customer has been a customer of the
bank by loss of customers.

Figure 5. Customer loss distribution of customer's account balance amount.

Figure 6 shows that the rate of leaving the bank by active customers (1) is
lower. In addition, it is seen that the number of active customers is higher.
It is seen that the loss of customers is approximately equal in every salary
range in Figure 7.
There are 60 people with a used bank product of 4 and they are all
abandoned customers. Figure 8 shows that customer loss is the highest
proportionally, after 3, customers using 1 and 2 bank products.
Analysis of Customer Churn in the Banking Industry Using Data Mining 65

Figure 6. Distribution of customer loss by customer activity.

Figure 7. Customer loss distribution of estimated salary.

Figure 8. Distribution of number of bank products used by the customer

by loss of customers.
66 Özge Doğuç

Data Preparation

The next stage of data analysis is data cleaning, making it usable for the model.
At this stage, the missing data in the data set is completed according to the
determined rules. Unnecessary, unwanted customer number, surname, etc.
variables are discarded. In this study, Knime was used, and the two variables
mentioned with the Column Filter node were cleaned from the dataset.
The Exited variable, which is our target (Class) variable, has been
converted into a string, that is, a categorical variable (string) with Number to
String node so that we can use it in the analysis. With the same node, the
variables of activity and credit card ownership were made categorical (string).
For a better analysis, new variables that are thought to be related to the
data set can be derived and added to the data set. In this study, the T/NOP
variable was produced, which indicates the ratio of the Usage time to the
Number of Products Used variables (Tenure/NumOfProducts). A Math
Formula node is used. Variables with continuous numerical values can be used
in the model by dividing them into categories. Rule Engine node is used.

Figure 9. Data preparation and data understanding phase.

In the data set, it was seen that the number of customers who did not leave
the bank was 7963, while the number of those who did was 2037. The big
difference between these two numbers is something that can make the model
difficult to learn. Therefore, the number of rows of the data set has been
increased by generating and adding data with the same statistical information,
without disturbing the structural features of the SMOTE node and the data set.
SMOTE does not reproduce existing values. Creates new values based on the
distance between existing values.
Analysis of Customer Churn in the Banking Industry Using Data Mining 67

Data preprocessing steps can be further multiplied. The stages in this

study are indicated. Figure 9 shows nodes in both data preparation and data
understanding on Knime.

Data Modeling

After the Dataset Understanding and Preparation stages comes the Model
creation stage. Various data mining algorithms that we will apply to the data
we have prepared are selected, the parameters are determined, and the model
is applied. When the model is ready, the accuracy value is calculated to
measure the model success. The model emerges because of processing the data
set with machine learning algorithms. Input information are hyperparameters
that may differ for each algorithm that we have given to the model beforehand.
The image shows its structure. In this study, Decision Tree, Random Forest,
and Artificial Neural Networks algorithms were used.

Decision Tree Algorithm

Decision Tree algorithm is one of the supervised machine learning algorithms

and is one of the most widely used classification algorithms. It is a predictive
model, and its structure is easier and more understandable than other
algorithms. The decision tree consists of nodes, the first node is the root node.
Each node asks a question about itself and according to the answer given, it
continues by connecting to another node as branches to the next nodes. When
it comes to an end, leaves representing a class are reached. The decision tree
consists of nested ifs.
The data set was divided into two as 70% for training and 30% for testing
with Partition node. With the training part, the Decision Tree algorithm model
was created. With the test data, the model estimated the target variable (class)
and compared it with the actual values. Thus, model success metrics and
Confusion Matrix were found. The matrix is shown in Table 4. The model was
built on the Knime platform, and the model image is shown in Figure 10 along
with the Random Forest algorithm model.
68 Özge Doğuç

Table 4. The Confusion Matrix of the Decision Tree model

Estimated One Zero

real value
One 1044 (TP) 756(FN)
Zero 579 (FP) 6621 (TN)

• True Positive (TP): True Positive => 1044

• False Negative (FN): False Negative => 756
• False Positive (FP): False Positive => 579
• True Negative (TN): True Negative => 6621

The Comparison Matrix given in Table 4, the model left the bank in the
test data set, that is, the model correctly predicted 1044 data with a value of 1
for the Churn variable. Didn't leave the bank The model correctly predicted
6621 pieces of data with a value of 0 as the Churn variable. Model 1 predicted
756 with a true value of 1 and model 1 with a true value of 0 and 579 with a
true value of 0. Based on this table, it is possible to easily measure the success
of the model with various values.

• Model Accuracy TP+TN/ Total: This metric shows us the proportion

of correct guesses among all predictions. The model gives its success.
In this study, it was 0.852. It should be increased by working on the
model.
• Precision of the Model (Precision) TP/(TP+FP): How many of those
estimated to have abandoned did. It was found to be 0.64 in this study.
• Sharpness of Model (Recall) TP/(TP+FN): Indicates how many of
the true abandoners were predicted. It is one of the most important
metrics. 0.58 was found in this study. It is expected to be higher. It
may be low due to the imbalance of 0 and 1 in the data set. It is
necessary to increase the model by making improvements.
• Misclassification Rate (FP+FN)/Total: Error rate of the model. 14.83
was found.
• F1 score 2*Precision*Recall/(Precision*Recall): It is a hybrid value
of the two weighted averages of the Sensitivity and Sharpness
metrics. It is beneficial to evaluate the two criteria together. 0.61 was
found.
Analysis of Customer Churn in the Banking Industry Using Data Mining 69

• Cohen's kappa coefficient of the model (Cohen's Kappa): It shows the

consistency and agreement between the categorical answer options.
0.519 found in this study

The Random Forest Algorithm

Random Forest is one of the supervised machine learning algorithms and is a

powerful tool for classification problems. The random forest algorithm allows
us to generate multiple trees. It is based on decision trees. It reduces the
problem of over-learning and overfitting, which is one of the biggest problems
of random forest decision trees. Each tree consists of the nodes in the decision
trees, the branches that connect them, and the leaves, which are the result. The
random forest provides a wider variety of modeling possibilities than the
decision tree algorithm. The main advantage of the random forest algorithm is
that it is very useful and easy to use. The number of hyperparameters we have
given is not much and is understandable.
It can give importance to variables. It indicates which variable is the root
node most frequently and shows the importance of this variable. In the bank
customer churn study, the variable that emerged the most root node was
repeated 28 times and became the Number of Product Usage variable. This
variable is the most important primary variable that is used most when making
the first distinction in trees. The most recurrent variable was Age. The decision
maker can use it to guide their decisions by disabling variables that are less
repetitive or giving importance to those that are highly repetitive.
In the Random Forest Decision Tree study, the model with the best
accuracy value was tried to be created by experimenting with hyper-
parameters. As a result of the trials, information gain was determined as the
split criterion, the limit number of levels of the trees was determined as 10,
and the minimum number of nodes (minimum node size) was determined as
4. The number of trees is the most important hyperparameter. While the high
number of trees increases the efficiency, it slows down the time. Therefore,
the number of models was determined as 100. The model was created with
these values and Table 5 gives the Confusion Matrix.
As in the Decision Tree, the data set was divided into two for 70% training
and 30% testing with Partition node. A Random Forest algorithm model was
created with the training part. With the test data, the model estimated the target
70 Özge Doğuç

variable (class) and compared it with the actual values. Thus, model success
metrics and Confusion Matrix were found. Table 3 shows the matrix.

Table 5. The complexity matrix of the Random Forest model

(Confusion Matrix)

Estimated One Zero

real value
One 762 (TP) 1038(FN)
Zero 183 (FP) 7017(TN)

• True Positive (TP): True Positive => 762

• False Negative (FN): False Negative => 1038
• False Positive (FP): False Positive => 183
• True Negative (TN): True Negative => 7017

Figure 10. Decision Tree and Random Forest algorithm models.

Analysis of Customer Churn in the Banking Industry Using Data Mining 71

The confusion matrix in Table 5 shows that the model left the bank in the
test data set, that is, 762 data with a value of 1 for the Churn variable were
correctly predicted as model 1. In terms of not to leave the ban, the model
correctly predicted 7017 pieces of data with a value of 0 for the Churn variable.
An incorrect estimation has been made by estimating 1038 with a true value
of 1 and model 0 with a true value of 183 with the model 1.

• The Accuracy value of the model was found to be 0.864.

• The Precision value of the model was found to be 0.806.
• The Sharpness (Recall) value of the model was found to be 0.423.
• The Misclassification Rate of the model was found to be 13,567.
• F1 score: 0.555 was found.
• Cohen's kappa coefficient of the model (Cohen's Kappa): 0.484 was
found.

The model set up on the Knime platform is shown in Figure 10.

Artificial Neural Networks

Artificial Neural Networks is an algorithm inspired by neurons and synapses,

which are the working logic of the human brain. In this algorithm, connections
are formed between artificial nerve cells and cells. There is an input layer, an
output layer, and there are hidden layers between them. The parameters of how
many layers and neurons should be determined in advance. There is no definite
rule, the appropriate number is given by trial and error. In hidden layers there
are neurons in the layer. Input layer has input(x) and output layer has target
(class) variables which we call output(y). The relationship between them is
y=f(x)+Error. To reduce the error weight (weight)(W) information is found
and assigned. The purpose here is to perform the operation with the inputs by
performing an operation according to the importance of the inputs. This is the
activation function.
These elements are present in the structure of artificial neural networks,
inputs, weights, summation function, activation function and outputs. The
working logic is the formation of the learning algorithm by finding the binding
weights (w) of the input elements. Explaining the results of this algorithm is
more difficult than explaining the result of a linear model. Algorithm layer
image is given in Figure 11. (Çınar, 2018)
72 Özge Doğuç

Figure 11. Artificial Neural Network example.

In artificial neural networks, data is used by normalizing. All data types

must be numeric for the data to be normalized. Categorical (string) converted
all input (input data) variables to numeric type with Category to Number node.
The target (class) variable is left categorical (string). The data were normalized
using the normalizer node and the z-score. Artificial neural network algorithm
cannot be established without normalizing the data. With the partitioning
node, the data was divided into two for 70% training and 30% for testing. The
model was built with the training part, then the model was tested with the test
data. The Artificial Neural Network model made in Knime is given in Figure
12. Table 4 gives the Confusion Matrix.

Figure 12. Artificial Neural Network algorithm model.

Analysis of Customer Churn in the Banking Industry Using Data Mining 73

Table 6. Complexity matrix of Artificial Neural Network model

(Confusion Matrix)

Estimated One Zero

real value
One 889 (TP) 944 (FN)
Zero 303 (FP) 6864 (TN)

• True Positive (TP): True Positive => 889

• False Negative (FN): False Negative => 944
• False Positive (FP): False Positive => 303
• True Negative (TN): True Negative => 6864

When we look at the Confusion Matrix given in Table 6, the model left
the bank in the test data set, that is, the model correctly predicted 889 data with
a value of 1 for the Churn variable. The model did not leave the bank. The
model correctly predicted 6864 data with a value of 0. Model 0 with a true
value of 1, 944 with a true value of 0, and 303 with a true value of 0, predicted
model 1.

• The Accuracy value of the model was found to be 0.861.

• The Precision value of the model was found to be 0.746.
• The Sharpness (Recall) value of the model was found to be 0.485.
• The Misclassification Rate of the model was found to be 13,856.
• F1 score: 0.588 was found.

Cohen's kappa coefficient (Cohen's Kappa) of the model was found to be

0.509.

Conclusion

In the study, data mining models were created that predict which bank
customers will leave the bank and which will not. The data comprehension
phase was done with Python programming language, and then the model was
created on the Knime platform. The data set used was divided into two. While
the model was created with the training part, the accuracy values of the model
were found with the test part. Algorithms used in modeling; Decision Tree
74 Özge Doğuç

(Random Forest) and Artificial Neural Networks (Artificial Neural Networks).

The truth value comparison table of these algorithms is given in table 7.

Table 7. Comparing algorithm metrics

Accuracy Sharpness (Recall)

Decision Tree 0.852 0.581
Random Forest 0.864 0.423
Artificial Neural Networks (Artificial Neural Networks) 0.861 0.485

According to table 7, the Accuracy value is a random forest algorithm,

albeit with slight differences. Another important metric is the decision tree
algorithm with the highest Recall value. The values found do not indicate that
the success of the model is at a sufficient level. Model success rates can be
increased by making changes on the hyperparameters used in the models,
adding new variables to the data set, cleaning the extra variables, or making
other changes on the model.
Making bank churn analyses of the bank's customers enables us to see the
customers who are likely to leave the existing customers. This is very
important information in terms of Customer Relationship Management and
Marketing activities of the relevant bank. You should not think only for
banking. Customer loss analysis is valuable for telecommunications, mobile
as well as other gaming channels, mobile applications, e-commerce sites and
many other companies. While the customer churn action is certain for
companies with subscription logic such as banking and telecommunications,
the situation may be different for other sectors. For example, an e-commerce
site “in the last x months must create the data set by specifying certain
conditions such as “customers who have not made any transactions.” The same
may be true for banks. Determining the customer churn value according to the
average time between transactions can be attributed to different conditions
than net situations such as account closure and card cancellation.
The effort and cost to acquire a new customer are greater than the cost of
retaining an existing customer. For this reason, companies need to carry out
activities to prevent loss of customers and loyalty programs. Firms will focus
on a more limited customer group, knowing in advance the customers who are
likely to leave. This will provide both ease of operation and cost advantage.
The activities that the companies can do by making use of the estimation
results of the customer churn probability obtained in this study are listed
below.
Analysis of Customer Churn in the Banking Industry Using Data Mining 75

• Customer churn rates can be reduced by making various campaigns

for customers who are likely to leave the company. Customers'
reactions to studies on such customers can also be recorded in the
database, and new data sets can be created and new analyzes can be
created. Thus, the success rates of the activity emerge. It can be a
guide for future campaigns.
• Reminding the most prominent features or services of the company
to customers who are likely to leave the company can be done via
SMS, pop-up, notification, or any other method. Recommendation
Engines, which are data analysis algorithms, can be used to determine
which feature will be recommended or reminded to which customer.
As stated in the previous article, the customer's reactions should be
recorded in the database so that they can be analyzed in future studies.
• Companies should regularly analyze customer loss and the results
should be shared with the departments. With the automatic mailing
node in Knime used in this study, the authorities can see the people
who are likely to leave the company, for example, once every 4
months during the year.
• Intra-company strategies can be created by paying attention to the
priority variables found in the Random Forest algorithm.
• By taking the abandoned customers into a separate data set, their
common features can be analyzed. With different data analysis
algorithms, the reasons for customer loss can be revealed by
comparing the customers who do not leave and who do not.

References

Alaybeg, F., 2019. Data Mining Introduction, Methods and Methodologies. Medium.
https://furkanalaybeg.medium.com/veri-madencili%C4%9Fi-ve-y%C3%B6ntemleri-
d0e2fd238e44.
Çınar, U., 2018. Neural Networks and Application with R Program. https://www.
veribilimiokulu.com/yapay-sinir-aglari/.
Doğan, B., Erol B. and Buldu, A., 2014. Using the Association Rule for Customer
Relationship Management in the Insurance Industry. International Journal of
Advances in Engineering and Pure Sciences 3, 105-114.
Gulpinar, V., 2013. Customer Loss Analysis in Turkish Telecommunications Market with
the Help of Artificial Neural Networks and Social Network Analysis. Marmara
University Journal of Economics and Administrative Sciences 34 (1), 331-350.
76 Özge Doğuç

Gupta, D. and Kamilla, U., 2014. Cyber Banking in India: A Cross-Sectional Analysis
Using Structural Equation Model. IUP Journal of Bank Management 13 (2).
Gürsoy, U. Ş., 2010. Customer Churn Analysis in Telecommunication Sector. Journal of
Istanbul University Faculty of Business 39 (1), 35–49.
Hsieh, N. C., 2004. An Integrated Data Mining and Behavioral Scoring Model for
Analyzing Bank Customers. Expert Systems with Applications 27 (4), 623–633.
Ivanchenko, O., Mirgorodskaya, O. N., Baraulya, E. V. and Putilina T. I., 2019. Marketing
Relations and Communication Infrastructure Development in the Banking Sector
Based on Big Data Mining. International Journal of Economics and Business
Administration 7, 176-184.
Maryani, I. and Riana, D., 2017. Clustering and Profiling of Customers Using RFM for
Customer Relationship Management Recommendations. In 2017 5th International
Conference on Cyber and IT Service Management (CITSM), 1-6.
Mukherjee, A. and Nath, P., 2003. A Model of Trust in Online Relationship Banking.
International Journal of Bank Marketing 5, 5-15.
Parviainen, P., Tihinen M., Kääriäinen J. and Teppola, S. 2017. Tackling the Digitalization
Challenge: How to Benefit from Digitalization in Practice. International Journal of
Information Systems and Project Management 5 (1), 63-77.
Payne, A., 2005. Handbook of CRM. Routledge.
Sassen, S., 2000. Globalization and Its Discontents: Essays on the New Mobility of People
and Money. New York: NewPress.
Savaşçı, I. and Tatlıdil R., 2006. The Effect of the Crm (Customer Relationship
Management) Strategy Applied by the Banks in the Credit Card Market on Customer
Loyalty. Aegean Journal of Academic Perspective 6 (1), 62-73.
Şimşek, H., 2018. Machine Learning Lessons 5a: Random Forest (Classification). Veri
Bilimi Türkiye. [Data Science Turkey].https://medium.com/data-science-en/machine-
%C3%B6%C4%9Frenmesi-desleri-5-bagging-ve-random-forest-2f803cf21e07.
Vafeiadis, T., Diamantaras, K. I., Sarigiannidis, G. and Chatzisavvas K. C., 2015. A
Comparison of Machine Learning Techniques for Customer Churn Prediction.
Simulation Modeling Practice and Theory 55, 1-9.
Wijaya, A. and Girsang, A. S., 2015. Use of Data Mining for Prediction of Customer
Loyalty. CommIT (Communication and Information Technology) Journal 10 (1), 41-
47.
Chapter 5

The Crowdsourcing Concept-Based

Data Mining Approach Applied
in Prosumer Microgrids

B. C. Neagu*, PhD, M. Gavrilaș, PhD,

O. Ivanov, PhD and G. Grigoraș
Department of Power Engineering, “Gheorghe Asachi” Technical University,
Iasi, Romania

Abstract

Energy efficiency is a tool for saving money and resources, representing

a necessity for flexible adaptation to users' demands. Because the use of
electricity plays a fundamental role in modern lifestyles, users' load
characteristics must reflect the people's lives at work and leisure. The
future microgrids must ensure “smart” features like flexibility,
accessibility, reliability, and high power quality for all consumers.
Increased adoption of small-scale distributed energy sources (SSDES)
helps lead to the decarbonisation of microgrids. Presently, with more
prominent reason, society requires to ensure that the benefits of smart
electricity are conceivable to the total and that the benefit is given in a
clean and effective manner. Energy poverty is described as a lack of
access to clean and affordable energy, resulting in soaring energy costs.
The crowdsourcing concept, introduced by J. Surowiecki in 2005, is used
to mitigate energy scarcity. It can be a useful tool for allowing the crowd
to do community service within a specific geographic region. According
to Romania's Energy Regulation National Agency's Order No. 228
launched on December 28, 2018, the prosumers can sell the energy

*
Corresponding Author’s Email: [email protected].

In: The Future of Data Mining

Editor: Cem Ufuk Baytar
ISBN: 979-8-88697-250-4
© 2022 Nova Science Publishers, Inc.
78 B. C. Neagu, M. Gavrilaș, O. Ivanov et al.

produced SSDES on the free market. Many advanced trading models

have as the goal increasing the benefits of the peers that sell energy in
local community market. Many researchers considered this field
challenging and significant to be investigated. Knowledge Discovery in
Databases (KDD) or Data Mining (DM) is the effort to understand,
analyse, and use a high amount of the available data. The main aim of
this paper is to quantify the distortion effects and introduce a stringent
and comprehensive methodology integrating the distribution network
operator (DNO), prosumers, and consumers. Based on aforementioned
considerations, an efficient and robust data mining-based methodology
is proposed to identify the power system cost saved by the residential
consumers when prosumers act to maximize their profit, and the DNOs
act to maximize the benefit resulted from an optimal operation of the
electricity distribution networks.

Keywords: crowdsourcing, data mining, prosumer, microgrids

Introduction

Increasing energy demand and the threat of global warming lead to the
exploitation of additional and cleaner energy sources. A consequence of this
trend is the growing penetration of SSDES. These installations may now
belong to microgrid consumers, which become prosumers, i.e., both producers
and consumers of energy. Photovoltaic (PV) panels, in particular, had
significant growth in recent years, with incentives given by EU countries like
Romania (see Law no. 184/2018), motivating its adoption and turning it into
a business case. The support of bidirectional power flows resulting from the
transactive energy in the local communities, as well as the need to decrease
the power loss lead to changes of the microgrids, namely in Low Voltage (LV)
distribution, pointing in the direction of a more responsive and efficient Smart
Grids (SGs).
Energy efficiency is a tool for saving money and resources, representing
a necessity for flexible adaptation to users' demands. Because the use of
electricity plays a fundamental role in modern lifestyles; users' load
characteristics must reflect the people's lives at work and leisure. The volume
of available information and necessary for the operation, management,
planning, and security of the microgrids have gradually increased with
technological development, requiring the introduction of the calculation
technique and intelligent solutions, Piccinini et al., (2015). Even if the
The Crowdsourcing Concept-Based Data Mining Approach … 79

prosumers have high benefits, their behavior is intermittent, and, for this
reason, the DNOs must consider a comprehensive microgrid planning
strategy. Moreover, they should be able to function both connected and
independently (autonomously) to the grid Andoni et al., (2010). Developed for
communities, the microgrids integrate local electricity production to satisfy
the exact demand of the consumers. The active consumers and prosumers are
defined in the context of recent paradigms of energy independence, energy
policy, and distributive management.
The microgrids is seen as the instrument to perform a coordinate structure
for fair integration of SSDES, which will be an edifying challenge for DNOs
that will require another operation plan. The prosumers form the active cells
of the microgrids in which any cell can supply installations to the DNO to
optimize the capacity of grid hosting and hence SSRES implementation in a
profitable and efficient operating way Espe et al., (2018).
In Romania, energy efficiency is improved from the increasing electricity
amounts generated by small producers (prosumers) with installed power up to
27 kW. These amounts are traded directly to the provider with which a
bilateral contract has already been signed Neagu et al., (2020a).
The prosumer's behavior must be a balanced one to store power surplus
when low demand spells and deliver it during the demand increases
Diahovchenko et al., (2020). An up-to-date report published by the European
Union - Smart Grids Task Force, Final report (2020), adds an explicit direction
to build a smart meter (SM) roadmap to satisfy the necessities of the future
energy markets through a modular and resilient structure for the metering
architecture. Thus, the SM should ensure helpful information on the shape of
the prosumers' load Chicco et al., (2020).
The proliferation of SSDES changes both the operating conditions and
management requirements of the microgrids, which must now integrate new
technologies and procedures. In recent years, the SG concept has been
implemented in several scales and initiatives in Romania, with much of the
hardware already tested and approved, albeit in isolation. In this context, an
assumption of a framework or unifying architecture of interfaces and protocols
based on norms and standards is essential. Thus, a reference architecture for
the data exchange between devices and electrical systems must be defined,
allowing the products, services, protocols, and interfaces to interact Kazmi et
al., (2017); Deng et al., (2017). Figure 1 illustrates the transition from classical
electricity grids to microgrids.
80 B. C. Neagu, M. Gavrilaș, O. Ivanov et al.

Figure 1. The transition from classical grids to microgrids.

Figure 2. The number of prosumers from each DSOs of Romania.

The increasing number of prosumers with various distributed energy

recourses promotes the peer-to-peer (P2P) energy transaction in the smart grid
for less cost, more flexibility, lower carbon footprints, and higher reliability.
Recent research on the prosumers’ behavior in the electric distribution systems
has increased in recent years. A distributed privacy-preserving P2P energy
transaction approach has been proposed to minimize the overall objective of
renewable generation curtailment penalty, adjustment cost, and operation cost
The Crowdsourcing Concept-Based Data Mining Approach … 81

while satisfying linearized distribution network power flow, power line

thermal, and voltage limit constraints. In the last two years, the prosumers'
integration in the microgrids has become an actual problem in Romania, with
the national government encouraging this process through incentives. For
example, Figure 2 shows the number of the over 3500 prosumers connected
in low voltage distribution networks for all Romanian DNOs Neagu et al.,
(2020b). In this context, the prosumers' behavior becomes a pressing issue for
aggregators, providers, and DNOs.
The system for promoting the production of renewable energy is
established by Law 184/2018. It represents an important step in the regulation
of prosumers’ status in Romania. This legislative framework provides for a
series of advantages, as follows (in accordance with the 228/2018 Order of the
Romanian National Regulatory Authority for Energy):

• the scheme is applied to prosumers owning renewable energy

production units with an installed capacity of not more than 27 kW
per consumption place in individual households, residential blocks,
residential, commercial or industrial areas;
• the electricity distribution operators must connect prosumers
following the specific regulations issued by the regulatory authority
in this respect;
• prosumers have the possibility of selling electricity to suppliers with
whom they have concluded electricity supply contracts at a price
equaling the weighted average price recorded on the day-ahead
market in the previous year; suppliers having an agreement with
prosumers must take over the energy at the former’s request;
• exemption of prosumers from the payment of excise duties for the
amount of electricity produced from renewable sources for self-
consumption, and the excess production sold to suppliers;
• exemption of prosumers as natural persons from the obligation of
purchasing green certificates annually and quarterly for the electricity
produced and used for own final consumption, other than own
technological consumption of power plants;
• benefiting from the regularisation service between the value of
electricity delivered and the value of electricity consumed in the grid
by the electricity suppliers with whom they have concluded electricity
supply contracts the service.
82 B. C. Neagu, M. Gavrilaș, O. Ivanov et al.

According to the above regulations, the electricity suppliers bound by

contracts with prosumers request to buy the electricity at the weighted average
day-ahead market price from the previous year. Thus, the prosumer can sell
on the market its electricity surplus, while the advantage for the supplier is the
exemption from the payment of the distribution network tariff. The trading
system offers a basic solution, having limitations regarding the options for
both parts (prosumers want to sell and consumers to buy electricity at lower
prices). It does not account for differences in generation costs and installed
capacity by not allowing prosumers to set custom selling prices. The incentive
of increasing local generation is not present. Consumers cannot buy electricity
directly from the prosumers, thus not having the freedom to choose specific
prosumers for trading Neagu et al., (2020c).
Many advanced trading models have as the goal increasing the benefits of
the peers that trade electricity in the local market of microgrids (LMM). A
comprehensive methodology for optimal integration of prosumers to minimize
active power losses in microgrids is proposed in the chapter. For a deeper
analysis regarding the influence of the prosumers on the microgrid power
losses, the 24-hour real load consumption and generation profiles are obtained
using a data mining process.

Crowdsourcing Energy System

The mitigation of energy poverty Neagu et al., (2020d) can be achieved using
crowdsourcing, a concept first introduced in 2005 by Surowiecki et al., (2005).
Crowdsourcing, Maxim et al., (2016), represents an emerging trend that
integrates contributions from users and the collective wisdom of the crowd,
Howe et al., (2006). It is very important to build a service-based approach to
make these crowdsourced sensor cloud data available. It can also be an
effective means to enable the crowd to provide a service-sharing community
within a geographical area by using their smartphones, Alt et al., (2010). Users
can take advantage of services from their neighborhood users through this
crowdsourced service community. Since the crowd (i.e., service providers) is
mobile, the availability of crowdsourced services to users is limited to its
spatio-temporal adjacency, i.e., both service providers and users should be
within a spatial region at a particular time.
A key issue is selecting and composing services from such a large number
of everchanging crowdsourced sensor cloud services to fulfill users’
requirements in a real-time fashion and based on spatio-temporal features. As
The Crowdsourcing Concept-Based Data Mining Approach … 83

a result, new spatio-temporal service selection and composition technologies

are key approaches to leverage spatio-temporal crowdsourcing as a service
provisioning platform.
A crowdsourced energy system (see Figure 3) includes a plurality of
distributed energy resources managed by crowdsources of the system, a power
network to which the distributed energy resources are connected, and a system
operator that manages energy trading transactions and energy delivery within
the system, the system operator operating at least one computing device
configured to obtain day-ahead peer-to-peer energy trading transaction
requests from crowdsources for energy to be delivered from the distributed
energy resources, estimate day-ahead energy load and solar forecasts,
determine optimal power flow for the delivery of energy, and schedule
delivery of energy from the distributed energy resources across the power
network based upon the energy trading transaction requests, the estimated
forecasts, and the determined optimal power flow.

Figure 3. A particular model for crowdsourcing energy system.

It can be an effective means to enable the crowd to provide community

service within a geographical area. In recent years, billions of dollars have
been invested in the research on blockchain technology to make the most of
its potential and understand how appropriate it is in the different economic
84 B. C. Neagu, M. Gavrilaș, O. Ivanov et al.

domains Mika et al., (2020). Nevertheless, not all fields are fully ready to
assimilate blockchain technology. The current technological opportunities
must be analyzed in each particular case together with the challenges that the
end-users face and with how a new decentralized architecture could create
value for them Wang et al., (2019).
Fortunately, the electricity industry is a suitable candidate for innovation
through blockchain technology. It integrates a complex supply chain with
needs to increase transparency and improved data management. In addition, it
contains a highly transactional trading market that would benefit from instant
settlement Unguru, (2018). The clarity and immutability of the blockchain can
empower end-users of this industry and consumers.

The Problem Formulation

The Prosumer Profiling Using Data Mining Method

The innovations in information technology have made it possible to acquire

and store large data amounts. Many activity fields, including the electricity,
are becoming increasingly dependent on data collection, storage, and
processing. However, the excess of data encounters difficulties in finding the
features that correspond to a specific objective. Due to the rapid increase of
the number and size of databases, it is necessary to examine and to propose
techniques for the automatic extraction of knowledge from these large size
databases. Many researchers considered this field challenging and significant
to be investigated. Knowledge Discovery in Databases (KDD) or Data Mining
(DM) is the effort to understand, analyse, and use a high amount of the
available data. Thus, a new research direction emerged that supported the
analysis of information extracted from existing data in databases, called data
mining. Generally, Data Mining represents a data analysis process to extract
useful information used to optimize performance indicators. There are
situations when a DNO has a large-size database, and a human operator could
not process it efficiently. Through applying a Data Mining algorithm, the
“hidden” features are discovered easily. These techniques can understand
patterns and apply them in feature selection using proper algorithms. The
features extracted from the databases are used in predictive models,
understanding the relationships between records or contents associated with
the database Neagu et al., (2020b).
The Crowdsourcing Concept-Based Data Mining Approach … 85

The research on the prosumers’ behaviour in the microgrids has increased

in the last years mainly due to the uncertain nature of the active power
injection into the network. If the difference between the generation and the
consumption of the prosumers is unknown, it can cause undesirable effects on
the optimal operation and planning of the microgrids. An efficient and robust
methodology based on the mining data techniques was proposed in the paper
to identify the behavioural characteristics of the prosumers from the
microgrids. There are three categories where the Data mining techniques can
be grouped, depending on the type of problem:

• Classification and regression. It represents the most common

category of applications, consisting of building the forecasting
models belonging to a set of classes (classification) or regression
values). From this category, the following approaches can be applied
with success: decision trees, the Bayes technique, neural networks,
and K-Nearest Neighbours.
• The analysis of associations and successions. This category generates
descriptive models that highlight correlation rules between the
attributes of a data set.
• Cluster analysis. It obtains the groups with similar entities or
highlights the entities that differ substantially from a group.

Figure 4 presents the main steps of Data Mining-based analysis. The

following operations take place inside each step Neagu et al., (2020b):

Figure 4. The components of the distribution service quality.

.

86 B. C. Neagu, M. Gavrilaș, O. Ivanov et al.

Data cleaning supposes a removal of irrelevant or “atypical” data;

Data integration considers the data sources integrated into the Data
Mining process. A common trend in information applications is to
perform data cleaning and integration as a pre-processing step, in
which the storage of resulting data is done in a data warehouse;
Data selection. The significant features are identified to be used in the
next step;
Data Normalization. Data are transformed or consolidated into forms
suitable for the data mining process;
Data Mining. It represents the essential process in which intelligent
methods are applied to extract patterns from data. K-means clustering
method was chosen in the power quality analysis being most used in
the scientific and industrial applications;
Pattern Evaluation. This step is necessary to identify the knowledge based
on particular measures of their degree of interest;
Knowledge presentation. Visualization and presentation techniques are
used to present to the Decision-Maker the discovered knowledge.

Optimal Allocation of Prosumers in Local Microgrids

The solution for the optimal location of small-scale renewable energy sources
(SSRES) uses a simulation algorithm based on iterative computation of steady
state, in order to minimize losses from the microgrid. The main objective is
the optimal location of generation sources minimize energy losses in a
microgrid. In essence, for a radial distribution network with N nodes, knowing
its structure, the optimization variables refer to the capacity of the distributed
generators known a priori, located in the nodes of the microgrid. In a first
phase, all the generation sources are located in each node of the microgrid.
The optimization variables are included in the column vector x whose length
is equal to the number of nodes, x = [x1, x2,…, xN]t where t is used for transpose
the vector. The inputs xi, (I = 1,…,N), represent a discrete value of the of
capacity size of the SSRES connected in a random node i (zero if the SSRES
is not present).
The optimization model considers the variation of the SSRES generated
power and the active and reactive load consumed. The considered analysis
timescale is divided into H time intervals Δth, for any h = 1,…,H. Neagu et al.,
(2019).
The Crowdsourcing Concept-Based Data Mining Approach … 87

The objective function is to minimize the energy losses indicated by ΔPh(d)

- the active power losses on each branch d = 1, ..., D in the time interval h =
1,…, H:

H D
min = F ( x ) =  P( ) ( x ) t
h =1 d =1
h
d
h (1)

However, a high degree of prosumers penetration can have a considerable

impact on power flows, increasing the voltage and power losses. Equality
restrictions are given by the equations of power flows; for any node i = 1,…,
N and for each time interval (h = 1,…, H), the equation is:

Ph(i ) + jQh(i ) = Uh(i ) Ih(i )* (2)

The inequality restrictions complete the optimization method with the

following:

1. Voltage variation limits:

U h( ,min
)
 U i , h  U h( ,max
)
i i
(3)

(i ) (i )
where U h,min and U h,max are the minimum and maximum effective value of
the voltage for bus I = 1,…, N at h=1,…, H.

2. Branch thermal limits: the effective value of the current over the brach
d = 1,…, D, at hour h= 1,…, H, denoted as I h( d ) , must be less than
(d )
the maximum admissible current for the line, I h,min :

I h( d )  I h( d,max
)
(4)

3. Generation limits: power generated by SSRES at bus I = 1,...,N in

hour h = 1,..., H, Ph(i, SSRES ) , is less than the maximum power allowed
(i , SSRES )
at bus, Ph,max :
88 B. C. Neagu, M. Gavrilaș, O. Ivanov et al.

Ph(i , SSRES )  Ph(,max

i , SSRES )
(5)

4. Constraints for reactive power in SSRES bus:

 Qh(
i , SSRES )
Qh(i,min
, SSRES )
 Qh( i,max
, SSRES )
(6)

( i , SSRES ) ( i , SSRES )
where Qh ,min and Qh,max are minimum/maximum allowable reactive
power of SSRES, and Qh(i , SSRES ) is the reactive power of SSRES in bus I =
1,…, N at hour h = 1,…, H.

The simulation search is used in the methodology by trying all

permutations of SSRES placements for the microgrid busses. The objective
function (1) is utilized in this example. Each alternative must be validated in
order to verify if the solution agrees to the restrictions. Validation of solutions
is required for each combination of SSRES locations through the following
steps: (i) power flow computation and elimination of solutions that result in at
least one restriction violation at one or more hours, (ii) objective function
evaluation for feasible solutions that do not violate any restrictions, and (iii) if
the new solution is better than the previous, update the best solution.

Results and Discussion

The proposed approach was implemented on a real LV microgrid from

Romania where the five PV panels (as SSRES) must be connected, and for
which the local generation profiles are considered in the 06:00 – 18:00
interval, values that can be checked in Neagu et al., (2020c). Its one-line
diagram is drawn in Figure 5. The μG has a simple structure, with 2 radial
feeders and 28 buses, and supplies a number of 27 residences, with one
consumer for each pole as is presented in Neagu et al., (2020a). The
optimization algorithm (1) – (5) is applied to minimize the power losses for a
summer working day. In the initial condition considered, the PV panels is not
connected – the daily losses are 149.25 kWh.
The optimal solution is presented in Figure 5. Thus, the consumers located
in buses 6, 7, 15, 21, 27 became prosumers trough the connection of the five
The Crowdsourcing Concept-Based Data Mining Approach … 89

PV panels. The buses where prosumers are present are depicted with blue light
on the house roof. The objective function for the optimal case (Scenario 0),
when all PV panels are connected to the microgrid, was about 96.92 kWh.
In order to carry out a deeper analysis regarding the impact of prosumers
operation on microgrid power losses, four scenarios were considered, namely:
Scenario 1 (only one prosumer is connected – five cases), Scenario 2 (two
prosumers are connected simultaneously – ten cases), Scenario 3 (three
prosumers are connected simultaneously - ten cases), and Scenario 4 (four
prosumers are connected simultaneously –five cases). All of the proposed
scenarios are reported to the initial case and the optimal scenario (Scenarios
0). For instance, the daily energy losses in the Scenario 1 are depicted in Figure
6.

Figure 5. The diagram of the tested LV microgrid.

The minimum value of the losses is obtained in the case when only the
prosumer no. 15 is in operation. The reduction is 20.23% compared to the
reference case. In the second scenario, there are ten cases as is shown in Figure
7. The minimum value of losses is obtained if the prosumers 7 and 15 are in
operation, with a reduction of the daily energy losses of 26.07% compared to
the reference case. For the last two Scenarios, the results are indicated in
Figures 8 and 9. In Scenario 3 the minimum value of losses is obtained if the
prosumers 7, 15 and 21 are in operation on the network, lower by 30.71%
compared to the reference case, while in Scenario 4 the energy losses is
reduced to 34.37% (the Pros. 27 is not in operation), significantly close to the
optimal scenario (35.06%).
90 B. C. Neagu, M. Gavrilaș, O. Ivanov et al.

Figure 6. The microgrid daily active energy losses in Scenario 1.

Figure 7. The microgrid daily active energy losses in Scenario 2.

In the most favourable case (Scenario 0 - all prosumers are in operation),

the total amount of energy produced by the PV panels in daylight time (06.00-
18.00) is 201.03 kWh. If to this quantity the 52.33 kWh (the daily reduced
values trough active energy losses) is added, results in a decrease of the
amount of energy extracted from the distribution network of 253.36 kWh.
This value is found as technological losses of the network operator as
declared in the literature, Neagu et al., (2020d).
The Crowdsourcing Concept-Based Data Mining Approach … 91

Figure 8. The microgrid daily active energy losses in Scenario 3.

Figure 9. The microgrid daily active energy losses in Scenario 4.

Technically, in the microgrids, prosumers would inject the surplus locally,

and the consumers would draw power in the same manner upon the power
flow laws. The results emphasize that the DNOs win through optimal
electricity flows between the prosumers, and the consumers with high power
demand. More than that, the prosumers, located as optimal case at buses 6, 7,
15, 21, and 27, can sell their electricity surplus only to the stakeholders Neagu
et al., (2020d).
92 B. C. Neagu, M. Gavrilaș, O. Ivanov et al.

Conclusion

This research examined a particular approach for optimal placement of small-

scale renewable electricity sources for microgrids. Traditionally, local
generation helps reduce the network losses. The modern concept of low
voltage microgrids, characterized by bidirectional power flow at the level of
the same consumer connection allows major changes in the operating
techniques of passive (classic) networks that become active distribution
networks. At this point, can talk about new types of end user, who are
generally more aware and more demanding, more active and more involved.
The prosumers form the active cells of the microgrids in which any cell can
supply installations to the DNO to optimize the capacity of grid hosting and
hence SSRES implementation in a profitable and efficient operating way. In
this context, the chapter highlights the obvious capabilities and benefits of
integrating prosumers in local microgrids according to the new legislation in
our country, to which are added the technical advantages, reduction of energy
losses (35.06% for the optimal scenario) by changing the active and reactive
powers flow in the microgrid using the crowdsourcing concept and data
mining approach. Furthermore, these profiles characterize the prosumers’
behaviour from the month with the maximum injected active energy. Based
on these approaches the DNO can develop optimal strategies to improve the
decision-making process regarding mainly the voltage control and phase load
balancing measures. The authors work now on the improvement of this
methodology to integrate all prosumers’ categories having various generation
systems (biomass, wind, hydro, biogas, and geothermal).

References

Alt, F., Shirazi, A. S., Schmidt, A., Kramer, U., and Nawaz, Z., (2010). Location-based
crowdsourcing: extending crowdsourcing to the real world. 6th Nordic Conference on
Human-Computer Interaction: Extending Boundaries, 13–22.
Andoni, M., Robu, V., Flynn, D., Abram, S., Geach, D., Jenkins, D., McCallum, P., and
Peacock, A., (2010). Blockchain technology in the energy sector: A systematic review
of challenges and opportunities, Renewable and Sustainable Energy Reviews, 143-
174.
Chicco, G., Labate, D., Notaristefano, A., and Piglione, F., (2020). Unveil the Shape: Data
Analytics for Extracting Knowledge from Smart Meters, Energia Elettrica
Supplement Journal, 1-16.
The Crowdsourcing Concept-Based Data Mining Approach … 93

Deng, S., Huang, L., Taheri, J., Yin, J., Zhou, M., and Zomaya, A. Y., (2017). Mobility-
Aware Service Composition in Mobile Communities, IEEE Transactions on Systems,
Man, and Cybernetics: Systems, 555–568.
Diahovchenko, I., Kolcun, M., Čonka, Z., Savkiv, V., and Mykhailyshyn, R., (2020).
Progress and Challenges in Smart Grids: Distributed Generation, Smart Metering,
Energy Storage and Smart Loads. Iran Journal Science Technology Transmission
Electrical Engineering, 1-15.
Espe, E., Potdar, V., and Chang, E., (2018). Prosumer Communities and Relationships in
Smart Grids: A Literature Review, Evolution and Future Directions. Energies, 2528.
European Smart Grids Task Force - Expert Group 3, Demand Side Flexibility - Perceived
barriers and proposed recommendations, Final Report, Apr. 2019.
Howe, J., (2006). The rise of crowdsourcing. Wired magazine, 1–4.
Kazmi, S. A. A., Shahzad, M. K., Khan, A. Z., and Shin, D. R., (2017). Smart Distribution
Networks: A Review of Modern Distribution Concepts from a Planning Perspective.
Energies, 501.
Maxim, A., Mihai, C., Apostoaie, C.-M., Popescu, C., Istrate, C., and Bostan, I., (2016).
Implications and Measurement of Energy Poverty across the European Union.
Sustainability, 483.
Mika, B., and Goudz, A., (2020). Blockchain-technology in the energy industry: blockchain
as a driver of the energy revolution? With focus on the situation in Germany. Energy
Systems.
National Regulatory Authority for Energy. The 228 Order for the Approval of the Technical
Norm Technical Conditions for Connection to the Public Electrical Networks of the
Prosumers; National Regulatory Authority for Energy: Bucharest, Romania, 2018.
Neagu B. C., and Grigoras, G., (2020a). A Fair Load Sharing Approach Based on Microgrid
Clusters and Transactive Energy Concept, 12nd Int. Conf. Electr. Comp. and Artif.
Intell., Bucharest, Romania, 1-4.
Neagu B. C., and Grigoras, G., (2020b). A Data-Mining-Based Methodology to Identify
the Behavioural Characteristics of Prosumers within Active Distribution Networks,
International Symposium on Fundamentals of Electrical Engineering 2020 (ISFEE),
Bucharest, Romania, 1-4.
Neagu, B. C., Ivanov, O., Grigoras, G., and Gavrilas, M., (2020c). A New Vision on the
Prosumers Energy Surplus Trading Considering Smart Peer-to-Peer Contracts.
Mathematics, 235.
Neagu, B.-C., Ivanov, O., Grigoras, G., Gavrilas, M., and Istrate, D.-M. (2020d). New
Market Model with Social and Commercial Tiers for Improved Prosumer Trading in
Microgrids. Sustainability, 7265.
Neagu, B. C., Grigoras, G., Ivanov, O., (2019). An Efficient Peer-to-Peer Based Blokchain
Approach for Prosumers Energy Trading in Microgrids. International Conference on
Modern Power Systems (MPS), Cluj Napoca, Romania, 1-4.
Piccinini, E., Gregory, R. W., and Kolbe, L. M., (2015). Changes in the producer-consumer
relationship-towards digital transformation. Changes, 1634-1648.
Surowiecki, J., (2005). The Wisdom of Crowds; Anchor: San Diego, CA, USA.
The Romanian Parliament, “Law no. 184/2018 for approving the Government Emergency
Ordinance no. 24/2017 regarding the modification and updating of Law no. 220/2008
94 B. C. Neagu, M. Gavrilaș, O. Ivanov et al.

for determining the incentive system for producing energy from renewable energy
sources and the modification of other normative acts,” Official Gazette, Part I, No.
635/20.07.2018.
Unguru, M., (2018). Blockchain technology: opportunities for the energy sector,
EUROINFO, 53-58.
Wang, S., Taha, A. F., Wang, J., Kvaternik, K., and Hahn, A., (2019), Energy Crowd-
sourcing and Peer-to-Peer Energy Trading in Blockchain-Enabled Smart Grids. IEEE
Transactions on Systems, Man, and Cybernetics: Systems, 1612–1623.
Chapter 6

Active Learning

Jože M. Rožanec1,2,3,*, Blaž Fortuna2

and Dunja Mladenić1
1Laboratory of Artificial Intelligence, Jožef Stefan Institute, Ljubljana, Slovenia
2Qlector d.o.o., Ljubljana, Slovenia
3Jožef Stefan International Postgraduate School, Ljubljana, Slovenia

Abstract

Active learning is a sub-field of machine learning concerned with how

an active learner can make decisions to draw a limited amount of data
instances to minimize the generalization error. The increasing
digitalization and growing abundance of data make the active learning
approach more critical than ever, introducing new challenges across
various disciplines. In the chapter, we introduce the field of active
learning, recent advances in the field and provide an overview of relevant
use cases across several disciplines.

Keywords: active learning, manufacturing, robotics, learning from

demonstration, healthcare

Introduction

Artificial Intelligence is a field of science devoted to learning principles,

techniques, and their application to leverage computers and machines to
mimic intelligent beings’ problem-solving and decision-making capabilities.

*
Corresponding Author’s Email: [email protected].

In: The Future of Data Mining

Editor: Cem Ufuk Baytar
ISBN: 979-8-88697-250-4
© 2022 Nova Science Publishers, Inc.
96 Jože M. Rožanec, Blaž Fortuna and Dunja Mladenić

Machine Learning is a sub-field of Artificial Intelligence that focuses on

learning such behaviors directly from the data through algorithms. It is thus
important that (i) the data is informative towards the goal the model aims to
achieve, (ii) the algorithm was suitably chosen to learn how to achieve the
goal, and (iii) that the model can learn from the data to achieve the goal.
Assuming that (ii) and (iii) were properly resolved and remain constant, the
models’ performance will depend on how informative the data is: the more
informative the data is, the higher the amount of content that can be learned
(Seung et al., 1992). Most machine learning algorithms are passive in the sense
that they are applied on a randomly selected dataset. Active learning, on the
other hand, attempts to influence how relevant instances are selected and
therefore impact the machine learning models’ learning over time (Tong and
Chang, 2001).

Figure 1. Taxonomy of active learning approaches. Active learning approaches can

be divided at least regarding four criteria: (a) how data is generated, (b) how data is
processed, (c) the size of the queries posed to the oracle, and (d) the problem solved.

Active Learning approaches and techniques can be characterized at least

according to four criteria as shown in Figure 1: (a) how the data is generated,
(b) how data is processed, (c) how many instances are queried at a time, and
(d) what machine learning problem is being solved. These four characteristics
determine the active learning approach. Membership querying synthesis is an
active learning approach that attempts to generate synthetic data to be later
Active Learning 97

labeled by an oracle. The challenge with such an approach is to create

instances that resemble the real-world data while introducing characteristics
that maximize the model’s learning. Approaches to select instances face
different constraints whether the data is processed in a streaming fashion (an
immediate decision must be taken, and memory constraints apply) or the data
can be processed offline. Instance selection for persistent data can be classified
into two categories (Wu, 2019): population-based (the test distribution is
known, and we aim to find the optimal training input density to generate the
train samples) and pool-based (select among unlabeled instances for labeling).
According to the number of instances queried in each iteration, we can divide
the approaches into sequential active learning (single instance) or batch active
learning (multiple instances) (Cai et al., 2017). Finally, the active learning
approaches differ based on the type of problem they solve (e.g., classification,
regression, ranking, and others) and how (e.g., supervised vs. unsupervised)
(Kai et al., 2006; Li et al., 2020).
We find supervised machine learning among machine learning
approaches, where algorithms are used to learn the mapping between input
feature values and their expected outcome. In such settings, active learning is
based on the premises that (i) unlabeled data is abundant, (ii) data labeling is
expensive, and (iii) the models’ generalization error can be minimized by
carefully selecting new input instances with which the model is trained (Kruk
et al., 2017; Sugiyama and Kawanabe, 2013). The assumption (iii) is grounded
on the consideration that the model’s bias is small enough to be ignored and
that the generalization error can thus be minimized by selecting the data used
to train the model. To that end, strategies and criteria were developed to find
the best-unlabeled instances to achieve such a purpose. Such a goal can be
achieved by selecting them from existing data (in a batch or streaming setting)
or generating them. When an unlabeled data instance is obtained, the active
learner can query the target value by requesting it to an oracle. The oracle can
be anyone or anything (e.g., an application programming interface or a human
annotator) that can accurately estimate the target value given the input feature
values. While Active Learning can be applied to regression and classification
problems, the former has received wider attention in the scientific community.
A supervised Active Learning setting comprehends two distinctive
elements: (a) means to obtain some useful, unlabeled data instance (a.k.a.
query since we are asking for a label); and (b) an oracle (e.g., a human data
annotator) that can provide a label for the aforementioned data instance. The
resulting labeled data instance is then used to train the machine learning
model, either immediately (in a streaming scenario) or upon model retraining
98 Jože M. Rožanec, Blaž Fortuna and Dunja Mladenić

(in a batch scenario, after incorporating it into the labeled dataset). We depict
the aforementioned elements in Figure 2. Three aspects must be considered
when looking for the most valuable samples (Wu, 2019): informativeness
(contains rich information that would benefit the objective function),
representativeness (how many other samples are similar to it), and diversity
(the samples do not concentrate in a particular region, but rather are scattered
across the whole space).

Figure 2. Schematic diagram of active learning. The oracle provides the data directly
to the ML model in an online setting. In contrast, in an offline setting, such instances
are persisted into a dataset so that the model can leverage them when retrained. The
diagram holds for most cases. An exception could be an oracle being directly asked
for a particular kind of demonstration, and therefore data selection would not be
needed.

In this chapter, we focus mainly on supervised active learning for

classification and regression. In section 2, we describe query strategies, mainly
for supervised active learning problems, while in section 3, we describe how
active learning was applied in use cases related to manufacturing, medicine,
and cybersecurity. Finally, we provide our conclusions in section 4.

Query Strategies

A frequent real-world scenario is the availability of large amounts of unlabeled

data, from which samples are drawn to create a dataset for supervised learning
purposes. Multiple approaches have been devised to that end, some of which
we describe. A simple baseline for selecting instances is the Random Active
Learning (RAL), which proceeds in rounds and, in each round, randomly
samples unlabeled data instances, assuming the data follows a uniform
distribution. Cuong, Lee, and Ye (2014) studied greedy active learning criteria
in pool-based active learning in a Bayesian setting. They compared the
maximum entropy criterion (selects the sample with the highest entropy)
Active Learning 99

(Settles, 2010), the least confidence criterion (selects the sample whose most
probable label has the least probability value) (Lewis, 1995), and the Gibbs
error criterion (selects the sample with the largest Gibbs error, which measures
the expected error of a Gibbs classifier predicting the label by sampling the
current label distribution) (Nguyen et al., 2013). Furthermore, they introduced
two greedy algorithms (Hamming loss and F1 loss), which achieved the best
performance under worst-case and average-case scenarios.
One of the most popular approaches is the query-by-committee (QBC)
framework (Seung et al., 1992), based on the idea that given a committee of
machine learning models, the committee’s variance can be indirectly
measured by examining the disagreement between their predictions.
Therefore, candidate data instances among the pool of unlabeled data are
selected based on the committee’s disagreement. Furthermore, multiple
strategies were proposed to measure such disagreement; e.g., Dagan et al.
(1993) proposed measuring disagreement using the vote entropy, considering
only the final class label regardless of the prediction scores. A variation to the
approach mentioned above is to measure the vote entropy by considering the
prediction scores of each committee member instead of the predicted labels.
Furthermore, a more sophisticated approach would consider measuring the
confidence of the committee members by measuring their divergence to the
mean and taking into account the class distribution of each committee member
(Kee et al., 2018).
Another approach is to consider the disagreement margin, which measures
how unanimous the committee is about the given prediction. Abe and
Mamitsuka (1998) proposed measuring the disagreement margin by
considering the difference of votes between the first and second most
confident label estimates for a given instance. Weigl et al., (2016) reframed
the aforementioned approach by considering the prediction scores. Another
approach was proposed by Borisov et al., (2011), who determined the
disagreement based on the standard deviation of the weighted prediction
scores of each committee member. Mccallum (1998) modified the QBC
approach using Expectation-Maximization to model the density over the
unlabeled data instances and use it to weight the disagreement scores to select
the unlabeled instances of interest to submit to the oracle.
Gammelsæter (2015) explored a different approach to QBC for neural
networks by applying a dropout layer to a Multi-Layer Perceptron (MLP) to
form a committee. Nevertheless, adding only one sample per time created an
unbalanced weight among the training samples. This issue was addressed by
Ducoffe and Precioso (2015), who described a QBC technique to train a
100 Jože M. Rožanec, Blaž Fortuna and Dunja Mladenić

convolutional neural network (CNN), leveraging batchwise dropout to create

a committee of partial networks given a full network trained by
backpropagation. The approach achieved a fast convergence rate and similar
error rates by using only 30% of the data available compared to an architecture
trained without the active learning approach and on the whole dataset.
While most active learning approaches consider the instance selection and
the model’s learning as separate tasks, Zhu et al., (2019) proposed considered
optimizing both objectives at once. The authors proposed a novel method for
QBC active learning of linear models referred to as Robust Zero-Sum Game
(RZSG) framework. The authors formulated the sample selection as a robust
optimization problem to find the best weights for unlabeled data to minimize
the true risk, minimize the average loss (increases robustness to outliers) and
minimize the maximal loss (increases robustness to imbalanced data
distributions). They compared this approach against another five methods for
data selection: (a) random sampling (sampling uniformly at random), (b) a
support vector machine margin-based querying technique for incorporating a
diversity of selected samples (Brinker, 2003), (c) a variance-based
regularization with convex objectives approach, which provides a theoretical
performance of the estimator with fast convergence by automatically
balancing bias and variance (Duchi and Namkoong, 2019), (d) an approach
based on Fisher Information matrices to select a batch of unlabeled samples
while reducing the redundancy between them (Zhang and Oles, 2000), and
(e) Batch Mode active leaning with Discriminative and Representative
queries, which attempts to balance informative and representative samples
based on the number of labeled samples (Wang and Ye, 2015). The authors
showed that the RZSG method leads to superior results while robust to
imbalanced data distribution and outliers and avoids sampling bias.
Searching instances close to the decision boundary has been frequently
realized with Support Vector Machine (SVM) models. Cohn and Schohn
(2000) described a simple heuristic to enhance the generalization behavior of
SVMs on classification tasks. They searched data samples orthogonal to the
space spanned by the training set, providing the learner with information about
dimensions not yet explored. They also envisioned picking examples along
the dividing hyperplane, narrowing the existing margin, and improving the
confidence in the dimensions the model already has information about. In the
same line, Tong and Chang (2001) proposed using uncertainty sampling for
an SVM model, querying the unlabeled data instances whose projection is
closest to the SVM hyperplane. Martens, Baesens and Gestel (2009), on the
other side, developed the Active Learning-Based Approach (ALBA) that
Active Learning 101

extracts rules from an SVM model to gain insight into its rationale and then
uses active learning to enhance the discrimination rules close to the SVM
decision boundary. A different take on the problem was developed by Guo and
Greiner (2007), who proposed querying unlabeled data instances, selecting
those providing the maximum conditional mutual information given the
labeled data, and using an optimistic guess for the unlabeled instances. There
are two possible scenarios given the optimistic guess: either the guess was
correct or not.
When the guess is incorrect, the authors propose selecting the most
uncertain unlabeled data instance, which helps to clarify the model’s decision
boundary. In the same line, Beygelzimer, Dasgupta and Langford (2009)
introduced the Importance Weighted Active Learning (IWAL), advocating the
relevance of importance weights to ensure a correct sampling bias. The learner
uses a specific strategy to estimate a rejection probability threshold based on
the unlabeled data instance and history of unlabeled samples queried to the
oracle. The weight of the particular instance is determined as the inverse of
the rejection probability, and used only if such data instance is queried against
the oracle. Another method for weighting instances was developed by Ganti
and Gray (2012), who devised the Unbiased Pool Based Active Learning
(UPAL) algorithm for pool-based active learning. The algorithm proceeds in
rounds, putting a probability distribution over the entire pool to then sample a
point from the pool. The probability distribution in each round depends on the
active learner and is obtained by minimizing the importance of weighted risk
over the hypothesis space. The authors demonstrated that UPAL outperforms
the Batch Mode Active Learning (BMAL) algorithm Hoi et al., (2006) in terms
of discriminative power, scalability, and execution time.
Xu et al., (2003) developed an approach to take into account the diversity
of data samples. To that end, the authors performed representative sampling
by clustering unlabeled data instances close to the decision boundary of an
SVM model and then selecting the medoid instances, which are expected to
preserve the density distribution information of the whole cluster. A similar
approach was followed by Nguyen and Smeulders (2004), who proposed using
clustering over all unlabeled data and then finding candidate samples
considering unlabeled data instances close to the decision boundary and
medoids of the highest density clusters. The one that contributes most to the
current error is chosen from the initial set of candidates. The authors argue that
while such a choice does not guarantee to achieve the smallest future error, it
is highly probable that such a choice would lead to a significant error decrease.
102 Jože M. Rožanec, Blaž Fortuna and Dunja Mladenić

Yang, Hanneke and Carbonell (2013) explored the use of active learning
in a transfer learning setting, where target concepts are sampled from an
unknown distribution. Furthermore, the authors explored the benefits of
transfer learning to know whether to stop the active learning procedure (self-
verifying active learning). In their research, they contrasted their work with
research from Baxter (1997) regarding the role of the sample size and loss
observed in learning multiple tasks. Yang, Hanneke and Carbonell (2013)
consider that the results are related only to the number of tasks. In contrast,
the number of samples per task remains bounded given non-altruistic learners
and the fact that the desired number of labeled samples should never be
significantly larger than the number of samples required to solve the task. The
authors consider that transfer learning can be used in a setting where all
problems cannot be solved by a single individual but rather by a coalition of
cooperating individuals who are willing to share the data used to learn a
particular concept to help others to learn another task. Furthermore, they
demonstrate that transfer learning can lead to a sample complexity close to the
one achieved by algorithms that directly know the targets’ distribution. Shao
(2019) also addressed the problem of actively selecting informative instances
aided by transferred knowledge from related tasks. To that end, the authors
proposed using a diverse committee with members from both domains to
select the most informative instances based on the maximum vote entropy,
improve the classification accuracy and evaluate each member over multiple
iterations.
A whole different view of the active learning problem was developed by
Zhu, Ghahramani and Lafferty (2003), who envisioned combining semi-
supervised learning and Active Learning by transforming labeled and
unlabeled data into vertices of a weighted graph connected by edges indicating
the similarity between data instances. While the semi-supervised learning
problem was framed in terms of a Gaussian random field model on the graph,
active learning was used to query unlabeled data and minimize the expected
classification error of the semi-supervised model. Active learning was also
applied to graph-specific problems. Ma, Garnett and Schneider (2013)
developed a graph-specific criterion called Σ-optimality, querying the nodes
that minimize the sum of the elements in the predictive covariance. Ostapuk,
Yang and Cudré-Mauroux (2019) developed an active learning framework for
knowledge graphs that takes into account the model uncertainty and the
underlying structure of the knowledge graph to improve the sampling
effectiveness and incrementally train deep learning models. A different
approach was developed by Li, Yin and Chen (2021), who proposed a
Active Learning 103

SEmisupervised Adversarial active Learning (SEAL) framework on attributed

graphs. The framework learns two adversarial components: (a) a graph
embedding network to encode labeled and unlabeled nodes into a common
latent space and (b) a semi-supervised discriminator network that learns to
distinguish between labeled and unlabeled nodes. The most informative
instances are then selected based on a divergence score generated by the
discriminator and submitted to an oracle for labeling.
Active learning has been successfully applied to regression problems too.
Wu (2019) considers three criteria must be taken into account when selecting
unlabeled samples: informativeness (the information richness of the samples,
which positively affects the objective function), representativeness (the
density of samples close to a certain target sample), and diversity (ensures the
samples are scattered across the full input space). The authors compare four
techniques: QBC, Expected Model Change Maximization (EMCM) Cai,
Zhang and Zhou (2013), Enhanced Batch Mode Active Learning for
Regression (EBMALR) (Wu et al., 2016) and greedy sampling (GS) (Yu and
Kim 2010). EMCM uses a set of labeled examples to create a linear regression
model M1 and then uses bootstrap to build a set of linear regression models
whose predictions are compared against the ones of M1 for a set of unlabeled
instances. To select the unlabeled instances, the algorithm considers those with
the maximum value given the average of the difference between predictions
of the M1 model and the set of linear regressors. EBMALR simultaneously
considers the informativeness, representativeness, and diversity of samples,
enhancing the QBC and EMCM approaches. To that end, it uses k-means
clustering to initialize a set of representative and diverse samples and then
recurs to baseline techniques to select subsequent samples sequentially.
Finally, greedy sampling attempts to find the most informative sample based
on the geometric characteristics of the feature space, selecting new samples
that are located far away from the previously selected and labeled samples.
The abovementioned methods described different strategies to select
existing data suitable for streaming and batch processing. The membership
query synthesis active learning strategy considers no data is selected, but
rather a data instance is created and presented to the oracle. J.-J. Zhu and Bento
(2017) introduced the Generative Adversarial Active Learning (GAAL)
technique, which leverages Generative Adversarial Networks (GANs) to
generate informative instances based on a random sample of unlabeled
instances close to the decision boundary. Mahapatra et al. (2018) evolved this
concept by generating synthetic data with a conditional GAN, which learns to
create a specific instance leveraging additional data regarding the desired
104 Jože M. Rožanec, Blaž Fortuna and Dunja Mladenić

target label and therefore leading to faster convergence. Sinha, Ebrahimi, and
Darrell (2019) introduced the variational adversarial active learning, sampling
instances using an adversarially trained discriminator to predict whether the
instance is labeled or not based on the latent space of the variational auto-
encoder. A weakness of this approach is that it can end up selecting instances
that correspond to the same class, regardless of the proportion of labeled
samples per class. This issue was addressed by Laielli et al. (2020), who
developed a semi-supervised mini-max entropy-based active learning
algorithm that leverages uncertainty and diversity in an adversarial manner. A
similar approach was described by Liu et al. (2019), who proposed the Single-
Objective Generative Adversarial Active Learning (SO-GAAL) technique,
which generates outliers with informative potential based on the mini-max
game between a generator and a discriminator to assist a classifier on
describing a boundary that separates outliers from normal data.
The overview of query methods provided above is by no means
exhaustive. We refer the reader to the following surveys to learn more about
active learning in particular settings. The surveys by Fu, Zhu and Li (2013),
and Kumar and Gupta (2020) provide great insights about query strategies for
classification, regression, and clustering in a batch setting. To learn about
techniques applied with deep learning models, we recommend the survey by
Ren et al. (2021). Finally, the survey by Lughofer (2017) provides an in-depth
introduction to active learning in an online setting.

Use Case

Industry and Robotics

While there is currently a research void regarding the use of active learning in
the manufacturing domain (Meng et al., 2020), the increasing digitalization
makes such approaches ever more relevant. Active learning has been
recognized to alleviate the manual labeling workload. Such quality has been
exploited in the manufacturing setting too. Furthermore, active learning is
considered one of the pillars of human-machine collaboration, key to the
Industry 5.0 paradigm (Rožanec et al., 2022a).
It can be argued that the manufacturing process begins with acquiring the
raw material and components required to manufacture the product. Accurate
demand forecasting is key to avoiding distortions across the supply chain.
Active Learning 105

Machine learning has been demonstrated to provide accurate forecasts

(Rožanec et al., 2021). In addition, explainable artificial intelligence methods
provide insights into the reasons behind such a forecast to craft explanations
and enable responsible decision-making (Rožanec et al., 2022b; Rožanec et
al., 2022c). Such explanations can be enriched with data from external
sources, recommending relevant pieces of data. In this line, Zajec et al., (2021)
explored different active learning strategies to combine recommendations and
active learning when displaying entries in such explanations so that the entries
were accurate and interesting to the user while relevant to the model’s learning
process. In a similar line of research, Li et al., (2017) developed a machine
learning model to assess whether input components required for the
manufacturing of fiber optic components will be delivered on time. The
manufacturing firm had valuable information to guide decision-making to
mitigate such risk by accurately predicting delayed customer orders. The
authors used active learning to select and label the most valuable samples that
are believed to maximize the classifier’s performance.
In manufacturing, significant resources are invested in qualifying
processes and machines that conform to quality and productivity standards.
One of such qualification procedures can be using experimental methods to
find the response surface mapping to process parameters. Botcha et al., (2021)
proposed using QBC-based active learning to find the next best experimental
point to reduce the uncertainty of prediction of surface roughness over the
sample space and avoid testing all possible parameters.
Active learning has been successfully applied to many quality inspection
use cases. Garderen (2017) explored its use in the quality process of measuring
the local displacement between layers on a chip, given that the ability to limit
overlay is key to the quality of semiconductor chips manufactured at a
nanometer scale. Quality measurements are performed in a separate metrology
tool. Since they are difficult to perform and time-consuming, only a fraction
of the produced wafers are analyzed. The authors compared several methods
which would be suitable for a regression problem that enabled forecasting and
visualization. Among the considerations they raised are that active learning
skews the sampling distribution, the results may have limited generalization
since selection results can vary across datasets, and that the sampling bias that
has been effective for a particular model does not provide performance
guarantees for others. In the same industry, Shim et al., (2020) describes how
wafer maps provide key information to engineers to detect root causes of
failure in the semiconductor manufacturing process. While a SOTA machine
learning model was built to classify them, a considerable cost is required to
106 Jože M. Rožanec, Blaž Fortuna and Dunja Mladenić

gather enough labeled data samples. The authors described an active learning
approach to select the wafer maps to alleviate this issue, ensuring no effort is
invested into data samples that would not provide enough information to the
classifier to learn. Dai et al., (2018) described a use case regarding automatic
optical inspection for the recognition of solder joint defects in printed circuit
boards. Active learning is applied to enlarge an initially labeled dataset. The
authors developed an original method combining semi-supervised learning
and active learning. In particular, they performed k-means clustering on
labeled and unlabeled data and trained a classifier on labeled data. They later
used the predictions issued by the classifier to analyze which clusters had the
least skewed distributions when considering the predicted classes of the
unlabeled samples and sample unlabeled data from them. A different use case
for visual inspection was described by Rožanec et al., (2022d); Trajkova et al.,
(2021), where multiple active learning approaches were considered to reduce
the manual visual inspection and labeling effort when inspecting the quality
of printed logos on manufactured shavers.
Active learning addresses the problem of which data to request to the
oracle to maximize the learning of a given agent. This idea strongly connects
with research on how robots are taught from human demonstrations and how
a physical robot can improve its skills by asking for such demonstrations. For
example, Maeda et al., (2017) describes an active learning approach that
enables robots to decide whether they have the skill to deal with an unknown
task or must ask for demonstrations to learn movement primitives
incrementally. In the same line of research, Conkey and Hermans (2019)
explored the use of active learning to learn a library of probabilistic movement
primitives, while Koert et al., (2019) used active learning to rely on fewer
demonstrations while improving the generalization capabilities when learning
a specific set of tasks. For a more in-depth overview of the use of active
learning in robotics, we refer the reader to the survey by Taylor, Berrueta, and
Murphey (2021).

Healthcare

Artificial intelligence can bring virtually unlimited progress to the field of

healthcare. One of the problems to which it has been successfully applied is to
annotate and help determine whether a patient suffers from a certain illness or
not. Much research was performed in the field of radiomics to perform such
analysis based on medical images. One such example is work by Doyle et al.,
Active Learning 107

(2011), who used weak decision tree classifiers to determine carcinoma and
non-carcinoma regions on prostate histopathological images. The authors used
a QBC active learning to annotate the images with the highest discrepancy
between the classification trees in a bagging setting. Padmanabhan et al.,
(2014) used artificial intelligence to detect and analyze the chemical and
physical characteristics of human cells in the context of renal carcinoma
detection. The authors reported using active learning to quantify the
information contribution of unlabeled samples by estimating the determinant
of the Fisher information matrix. The samples expected to provide the greatest
contribution were selected for manual labeling. Das, Nair, and Peter (2020)
described the use of artificial intelligence to analyze histopathological tissue
sample images and determine variations in size and shape that correspond to
malignant lesions associated with breast cancer. Given that the annotation of
such images requires a considerable effort of domain experts and experienced
pathologists, the authors used Active Learning to select batches of instances
based on the kernelized Riemannian distance measures, such as the Jeffrey
and Stein divergences. Among the results, they report using only 20% of the
labeled data required to train a supervised classifier model without the active
learning setting. A different approach was developed by Doyle and
Madabhushi (2010), who used a consensus of ambiguity to identify images
considered ambiguous by multiple algorithms and prioritize them for manual
labeling. Active learning has been also successfully applied to image
classification models related to the detection of colorectal cancer (Zhao et al.,
2019), skin lesions (Shi et al., 2019) or other diseases, such as COVID-19 or
pneumonia (Wu et al., 2021; Nguyen et al., 2021).
Accurate diagnostics require the right level of medical expertise to assess
the evidence. Mu et al., (2021) reported using deep learning to map semi-
structured and unstructured text data from pathology synopses to provide
relevant semantic diagnostic labels. Such a system attempts to avoid
bottlenecks that result from the limited number of specialists available to
interpret the pathology synopses and provide relevant diagnostic information.
In addition, the authors used active learning to prioritize the labeling of
unlabeled samples considering those that are underrepresented in the dataset
or the ones that could enhance the models’ performance.
Once a disease is discovered, it is crucial to understand how critical it is
to ensure the best treatment possible for the patient while also meeting the
overall patient scheduling constraints. This problem was addressed by Nissim
et al., (2017), who developed the CAESAR-ALE (Classification Approach for
Extracting Severity Automatically from Electronic Health Records - Active
108 Jože M. Rožanec, Blaž Fortuna and Dunja Mladenić

Learning Enhancement) framework to determine the urgency of further

diagnoses and the seniority of personnel required to attend to such patients.
The authors reported using an SVM classifier to classify the clinical records
and leverage active learning to reduce the labeling effort between 48% to 64%
compared to a standard passive SVM model.
Active learning has also been successfully applied to guide clinical trials.
One such example is research done by Minsker et al., (2016), who explored
the usage of artificial intelligence to optimize the selection of patients for
clinical trials to learn the best individual treatment rules. The authors argue
that randomized clinical trials are expensive and usually not efficiently
designed to estimate the individual treatment rules. Therefore, they use Active
Learning to estimate which persons can be considered the most informative
patients and let such insights guide further recruitment over the ongoing
clinical trials. They exclude the patients on whom the treatment effects are
clearly observed and focus on those where further evidence is required,
reducing overall sample sizes and associated costs.
Artificial intelligence can also be applied to search for proper treatment
and cures. Jin et al., (2020) described how academic papers are retrieved and
analyzed to understand reported critical clinical evidence related to a given
treatment in populations characterized by a disease and gene mutation. They
use Active Learning to leverage expert annotations to enhance a ranking
model that considers scores from an information retrieval system, a pre-trained
BioBERT model, publication type, and citation counts. The BioBERT model
is used for each iteration to propose relevant unannotated query-document
instances and then finetuned given the experts’ input. A different use case was
proposed by Danziger et al., (2009), who used artificial intelligence to
determine which mutations result enables the production of proteins with a
particular function. In particular, they sought mutations in the tumor
suppressor protein p53 since p53 mutants are found in half of all human
cancers. Restoring the non-mutant p53 proteins in tumors leads to their
decrease. The authors developed a specific active learning strategy (Most
Informative Positive) tailored to biological problems to select functionally
active examples compatible with further exploration of combinatorial
mutagenesis. The above strategy enabled reducing the number of experiments
by a third without a significant classification performance decrease.
Active Learning 109

Cybersecurity

The increasing complexity of cyberspace security requires intelligence-driven

cybersecurity defense. Machine learning applications to the cybersecurity
domain range from the development and mining of knowledge graphs (to
provide intelligence to the security analyst) to detecting and preventing
cybersecurity attacks. While the construction of knowledge graphs based on
deep learning has achieved great success, it requires the annotation of large
corpora. The annotation effort can be alleviated through active learning. One
such approach is described by Xie et al., (2021), who combined active learning
and deep learning frameworks to recognize and annotate Chinese
cybersecurity entities in Chinese texts. To that end, they build upon the work
of active learning for entity recognition by Shen et al., (2017), and Chinese
cybersecurity entity recognition by Qin et al., (2019) and propose a selection
strategy with uncertainty and confidence while considering the lexicon. The
authors used a BERT pre-trained model and a residual dilation convolutional
neural network to learn entity context features and conditional random fields
for tag decoding. They complemented the model with an active learning
approach, selecting the highest uncertainty cases computed considering the
posterior probability. Given that the method tends to favor long sentences, the
authors computed the confidence of the label of the sentence in the decoding
process and the matching frequency of cybersecurity entities in a lexicon to
evaluate whether the sentence should be sent to manual revision.
The use of active learning to guide manual revision has also been
leveraged for other problems. Chung et al., (2020) performed research
regarding the monitoring of internal networks to prevent data exfiltrations.
The authors developed an active learning approach to complement anomaly
detection models with manual revision to deal with the many false alarms and
learn from them. To that end, they used a LightGBM classifier and queried
unlabeled samples based on their entropy. Nevertheless, the authors observed
that while the models’ overall performance increases over time, it can lead to
the fact that the model can improve its performance for certain classes while
degrading others with each retrain. Similarly, Dang (2020) described using
active learning in the context of an intrusion detection system to reduce the
labeling cost while maintaining the classification performance. They trained a
Naive Bayes algorithm and selected rare events detected by the Isolation
Forest, which measures the difference of an instance regarding the observed
distribution. The decision of whether to include the selected instance into the
train set is made based on the predicted performance impact in terms of AUC
110 Jože M. Rožanec, Blaž Fortuna and Dunja Mladenić

ROC change by an ad-hoc XGBoost model. Concept drift can make machine
learning models vulnerable to attackers who progressively distill attack traces
into models and make them hard to distinguish from the values change related
to the concept drift. Dey et al., (2020) explored using active learning to
leverage human expertise and maintain effective detection capabilities even in
the face of concept drift regarding behavioral models.
Active learning was also used to timely annotate data for other
cybersecurity-related problems. Das Bhattacharjee et al., (2017), who describe
the use of active learning to create a semi-supervised solution to annotate data
regarding Phishing categorization. Among the particular challenges of the
domain is the need for real-time detection (the user must be warned about the
potential danger before (s)he clicks on the URL) and the ability to generalize
beyond the blacklisted URLs. The authors used the uncertainty sampling
strategy to achieve greater clarity near the classifier’s decision boundaries and
increase the model’s performance.

Conclusion

The active learning sub-field of machine learning is concerned with finding

the right data instances that maximize a given algorithm’s learning goal.
Through this chapter, we introduced a taxonomy, showing how multiple
criteria constrain and influence the development and choice of active learning
methods. We then provided a high-level overview of query strategies to select
or create relevant data samples and described research on use cases from
manufacturing and robotics, healthcare, and cybersecurity.
The decreased cost of sensors and digital devices favors an increasing
digitalization of every aspect of our lives. The abundance of collected data and
increasing ease of access to it has enabled the development of machine
learning models and products addressing a wide range of needs. Furthermore,
it has refocused the attention of research back to the quality of data with a
particular concern on how it is helpful to solve the particular problem at hand.
We consider active learning to play a crucial role in the renewed attention
towards data-centric artificial intelligence, developing a new understanding of
how to acquire data that drives the best learning outcomes for the algorithms
and overall artificial-intelligence-powered solutions.
Active Learning 111

References

Abe, N., Mamitsuka, H., 1998. Query Learning Strategies Using Boosting and Bagging. In
ICML ’98: Proceedings of the Fifteenth International Conference on Machine
Learning, 1–9.
Baxter, J., 1997. A Bayesian/Information Theoretic Model of Learning to Learn via
Multiple Task Sampling. Machine Learning 28 (1), 7–39. https://doi.org/10.1023/A:
100 73 27622663.
Beygelzimer, A., Dasgupta, S., Langford, J., 2009. Importance Weighted Active Learning.
Proceedings of the 26th International Conference On Machine Learning, ICML 2009,
no. 2008, 49–56.
Bhattacharjee, S. D., Talukder, A., Al-Shaer, E., Doshi, P., 2017. Prioritized Active
Learning for Malicious URL Detection Using Weighted Text-Based Features. 2017
IEEE International Conference on Intelligence and Security Informatics: Security and
Big Data, ISI 2017, 107–12. https://doi.org/10.1109/ISI.2017.8004883.
Borisov, A., Tuv, E., Runger, G., 2011. Active Batch Learning with Stochastic Query-by-
Forest (SQBF). Workshop on Active Learning and Experimental Design 16, 59–69.
Botcha, B., Iquebal, A. S., Bukkapatnam, S. B., 2021. Efficient Manufacturing Processes
and Performance Qualification via Active Learning: Application to a Cylindrical
Plunge Grinding Platform. Procedia Manufacturing 53 (2020), 716–725. https://doi.
org/10.1016/j.promfg.2021.06.070.
Brinker, K,, 2003. Incorporating Diversity in Active Learning with Support Vector
Machines. In Proceedings of the Twentieth International Conference on Machine
Learning (ICML-2003). https://doi.org/10.1039/C19680000233.
Cai, W., Muhan Z., Ya Z., 2017. Batch Mode Active Learning for Regression with
Expected Model Change. IEEE Transactions on Neural Networks and Learning
Systems 28 (7), 1668–1681. https://doi.org/10.1109/TNNLS.2016.2542184.
Cai, W., Zhang, Y., Zhou, J., 2013. Maximizing Expected Model Change for Active
Learning in Regression. Proceedings - IEEE International Conference on Data
Mining, ICDM, 51–60. https://doi.org/10.1109/ICDM.2013.104.
Chung, M-H., Chignell, M., Wang, L., Jovicic, A., Raman, A., 2020. Interactive Machine
Learning for Data Exfiltration Detection: Active Learning with Human Expertise.
Conference Proceedings - IEEE International Conference on Systems, Man and
Cybernetics 2020-October: 280–287. https://doi.org/10.1109/SMC42975.2020.928
2831.
Conkey, A., Hermans, T., 2019. Active Learning of Probabilistic Movement Primitives.
IEEE-RAS International Conference on Humanoid Robots 2019-October, 425-432.
https://doi.org/10.1109/Humanoids43949.2019.9035026.
Cuong, N. V., Lee, W. S., Ye, N., 2014. Near-Optimal Adaptive Pool-Based Active
Learning with General Loss. Uncertainty in Artificial Intelligence - Proceedings of the
30th Conference, UAI 2014, 122–131.
Dagan, I., Engelson, S. P., and Gan, R., 1993. Committee-Based Sampling For Training
Probabilistic Classifiers. Science.
112 Jože M. Rožanec, Blaž Fortuna and Dunja Mladenić

Dai, W., Mujeeb, A., Erdt, M., Sourin, A., 2018. Towards Automatic Optical Inspection of
Soldering Defects. Proceedings - 2018 International Conference on Cyberworlds, CW
2018, 375–382. https://doi.org/10.1109/CW.2018.00074.
Dang, Q. V., 2020. Proceedings - 2020 RIVF International Conference on Computing and
Communication Technologies, RIVF 2020. In 2020 RIVF International Conference
on Computing and Communication Technologies (RIVF), 13.
Danziger, S. A., Baronio, R., Ho, L., Hall, L., Salmon, K., Hatfield, G. W., Kaiser, P.,
Lathrop, R. H., 2009. Predicting Positive P53 Cancer Rescue Regions Using Most
Informative Positive (MIP) Active Learning. PLoS Computational Biology 5 (9).
https://doi.org/10.1371/journal.pcbi.1000498.
Das, A., Nair, M.S., Peter, D.S., 2020. Batch Mode Active Learning on the Riemannian
Manifold for Automated Scoring of Nuclear Pleomorphism in Breast Cancer.
Artificial Intelligence in Medicine 103, 101805. https://doi.org/10.1016/j.artmed.
2020.101805.
Dey, A., Velay, M., Fauvelle, J-P., Navers, S., 2020. Adversarial vs Behavioural-Based
Defensive AI with Joint, Continual and Active Learning: Automated Evaluation of
Robustness to Deception, Poisoning and Concept Drift, 1–25. http://arxiv.org/abs/
2001.11821.
Doyle, S., Madabhushi, A., 2010. Consensus of ambiguity: theory and application of active
learning for biomedical image analysis. In IAPR International Conference on Pattern
Recognition in Bioinformatics, 313–324.
Doyle, S., Monaco, J., Feldman, M., Tomaszewski, J., Madabhushi, A., 2011. An Active
Learning Based Classification Strategy for the Minority Class Problem: Application
to Histopathology Annotation. BMC Bioinformatics 12, 1–14. https://doi.org/
10.1186/1471-2105-12-424.
Duchi, J., Namkoong H., 2019. Variance-Based Regularization with Convex Objectives.
Journal of Machine Learning Research 20 (3), 1–10.
Ducoffe, M., Precioso, F., 2015. QBDC: Query by Dropout Committee for Training Deep
Supervised Architecture, no. 2013, 1–10. http://arxiv.org/abs/1511.06412.
Fu, Y., Zhu, X., Li, B., 2013. A Survey on Instance Selection for Active Learning.
Knowledge and Information Systems 35 (2), 249–283. https://doi.org/10.1007/s10115-
012-0507-8.
Gammelsæter, M., 2015. A Committee of One - Using Dropout for Active Learning in Deep
Networks | Semantic Scholar. 2015.
Ganti, R., Gray, A., 2012. UPAL: Unbiased Pool Based Active Learning. Journal of
Machine Learning Research 22 (422), 422–431.
Garderen, K. V., 2017. Active Learning for Overlay Prediction in Semi-Conductor
Manufacturing 6 (0), 1–18.
Guo, Y., Greiner, R., 2007. Optimistic Active Learning Using Mutual Information. IJCAI
International Joint Conference on Artificial Intelligence, 823–829.
Hoi, S. C. H., Jin, R., Zhu, J., Lyu, M. R., 2006. Batch Mode Active Learning and Its
Application to Medical Image Classification. ACM International Conference
Proceeding Series 148, 417–424. https://doi.org/10.1145/1143844.1143897.
Jin, Q., Tan, C., Chen, M., Yan, M., Huang, S., Zhang, N., Liu, X., 2020. Aliababa DAMO
academy at TREC precision medicine 2020: State-of-the-art evidence retriever for
Active Learning 113

precision medicine with expert-in-the-loop active learning. In Proceedings of the

Twenty-Ninth Text REtrieval Conference, TREC, 16–20.
Kai, Y., Bi, J., Volker, T., 2006. Active Learning via Transductive Experimental Design.
ACM International Conference Proceeding Series 148, 1081–1088. https://doi.org/
10.1145/1143844.1143980.
Kee, S., del Castillo, E., Runger, G., 2018. Query-by-Committee Improvement with
Diversity and Density in Batch Active Learning. Information Sciences 454, 401–418.
https://doi.org/10.1016/j.ins.2018.05.014.
Koert, D., Pajarinen, J., Schotschneider, A., Trick, S., Rothkopf, C., Peters J., 2019.
Learning Intention Aware Online Adaptation of Movement Primitives, in IEEE
Robotics and Automation Letters, 4 (4), 3719-3726.
Kruk, M. E., Emilia, J. L., Asaf, B., Melani, C., Karen, C., Mickey, C., Fadi, E. J., 2017.
Building Resilient Health Systems: A Proposal for a Resilience Index. BMJ (Online).
https://doi.org/10.1136/bmj.j2323.
Kumar, P., Gupta, A., 2020. Active Learning Query Strategies for Classification,
Regression, and Clustering: A Survey. Journal of Computer Science and Technology
35. https://doi.org/10.1007/s11390-020-9487-4.
Laielli, M., Zhu, S., Darrell. T., 2020. Minimax active learning. arXiv preprint
arXiv:2012.10467.
Lewis, D. D., 1995. A Sequential Algorithm for Training Text Classifiers. ACM SIGIR
Forum 29 (2), 13–19. https://doi.org/10.1145/219587.219592.
Li, C., Handong, M., Zhao, K., Ye, Y., Xiao, Y.Z., Guoren, W., 2020. On Deep
Unsupervised Active Learning. IJCAI International Joint Conference on Artificial
Intelligence 2021-Janua, 2626–2632. https://doi.org/10.24963/ijcai.2020/364.
Li, Q., Wei, F., Zhou, S., 2017. Early Warning Systems for Multi-Variety and Small Batch
Manufacturing Based on Active Learning. Journal of Intelligent and Fuzzy Systems
33 (5), 2945–2952. https://doi.org/10.3233/JIFS-169345.
Li, Y., Yin, J., Chen, L., 2021. SEAL: Semisupervised Adversarial Active Learning on
Attributed Graphs. IEEE Transactions on Neural Networks and Learning Systems 32
(7), 3136–3147. https://doi.org/10.1109/TNNLS.2020.3009682.
Liu, Y., Li, Z., Zhou, C., Jiang, Y., Sun, J., Wang, M., He, X., 2019. Generative Adversarial
Active Learning for Unsupervised Outlier Detection. IEEE Transactions on
Knowledge and Data Engineering 32 (8), 1517–1528. https://doi.org/10.1109/TKDE.
2019.2905606.
Lughofer, E., 2017. On-Line Active Learning: A New Paradigm to Improve Practical
Useability of Data Stream Modeling Methods. Information Sciences 415–416, 356–
376. https://doi.org/10.1016/j.ins.2017.06.038.
Ma, Y., Garnett, R., Schneider, J., 2013. ∑-Optimality for Active Learning on Gaussian
Random Fields. Advances in Neural Information Processing Systems, 1–9.
Maeda, G., Ewerton, M., Osa, T., Busch, B., Peters, J., 2017. Active Incremental Learning
of Robot Movement Primitives. Conference on Robot Learning (CoRL), no. CoRL,
37–46.
Mahapatra, D., Bozorgtabar, B., Thiran, J. P., Reyes, M., 2018. Efficient Active Learning
for Image Classification and Segmentation Using a Sample Selection and Conditional
Generative Adversarial Network. Lecture Notes in Computer Science (Including
114 Jože M. Rožanec, Blaž Fortuna and Dunja Mladenić

Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

11071 LNCS, 580–588. https://doi.org/10.1007/978-3-030-00934-2_65.
Martens, D., Baesens, B., Gestel, T. V., 2009. Decompositional Rule Extraction from
Support Vector Machines by Active Learning. IEEE Transactions on Knowledge and
Data Engineering 21 (2), 178–191. https://doi.org/10.1109/TKDE.2008.131.
Mccallum, A., 1998. Employing EM in Pool-Based Active Learning for Text Classification
1 Introduction. ICML '98: Proceedings of the Fifteenth International Conference on
Machine Learning.
Meng, L., McWilliams, B., Jarosinski, W., Park, H. Y., Jung, Y. G., Lee, J., Zhang, J., 2020.
Machine Learning in Additive Manufacturing: A Review. Jom 72 (6), 2363–2377.
https://doi.org/10.1007/s11837-020-04155-y.
Minsker, S., Zhao, Y. Q., Cheng, G., 2016. Active Clinical Trials for Personalized
Medicine. J Am Stat Assoc. 176 (3), 139–148. https://doi.org/10.1080/01621459.2015.
1066682.Active.
Mu, Y., Tizhoosh, H. R., Tayebi, R. M., Ross, C., Sur, M., Leber, B., Campbell, C. J. V.,
2021. A BERT Model Generates Diagnostically Relevant Semantic Embeddings from
Pathology Synopses with Active Learning. Communications Medicine 1 (1).
https://doi.org/10.1038/s43856-021-00008-0.
Nguyen, C. D. T., Huynh, M. T., Tran, M. Q., Nguyen, N. H., Jain, M., Van, D. N., Vo,
T.D., Bui, T. H., Truong, S. Q. H., 2021. GOAL: Gist-Set Online Active Learning for
Efficient Chest X-Ray Image Annotation. Proceedings of Machine Learning Research
143, 545–553.
Nguyen, H. T., Smeulders, A., 2004. Active Learning Using Pre-Clustering. Proceedings,
Twenty-First International Conference on Machine Learning, ICML 2004, 623–630.
https://doi.org/10.1145/1015330.1015349.
Nguyen, V. C., Wee, Sun, L., Nan, Y., Kian, M. A. C., Hai, L. C., 2013. Active Learning
for Probabilistic Hypotheses Using the Maximum Gibbs Error Criterion. Advances in
Neural Information Processing Systems, 1–9.
Nissim, N., Shahar, Y., Elovici, Y., Hripcsak, G., Moskovitch, R., 2017. Inter-Labeler and
Intra-Labeler Variability of Condition Severity Classification Models Using Active
and Passive Learning Methods. Artificial Intelligence in Medicine 81, 12–32.
https://doi.org/10.1016/j.artmed.2017.03.003.
Ostapuk, N., Yang, J., Cudré-Mauroux, P., 2019. ActiveLink: Deep Active Learning for
Link Prediction in Knowledge Graphs. The Web Conference 2019 - Proceedings of
the World Wide Web Conference, WWW 2019, 1398–1408. https://doi.org/10.1145/
3308558.3313620.
Padmanabhan, R. K., Somasundar, V. H., Griffith, S. D., Zhu, J., Samoyedny, D., Tan, K.
S., et al., 2014. An Active Learning Approach for Rapid Characterization of
Endothelial Cells in Human Tumors. PLoS ONE 9 (3). https://doi.org/10.1371/journal.
pone.0090495.
Qin, Y., Shen, G. W., Zhao, W. B., Chen, Y. P., Yu, M., Jin, X., 2019. A Network Security
Entity Recognition Method Based on Feature Template and CNN-BiLSTM-CRF.
Frontiers of Information Technology and Electronic Engineering 20 (6), 872–884.
https://doi.org/10.1631/FITEE.1800520.
Active Learning 115

Ren, P., Xiao, Y., Chang, Huang., P. Y., Li,, Z., Gupta, B. B., Chen, X., Wang, X., 2021.
A Survey of Deep Active Learning. ACM Computing Surveys 54 (9).
https://doi.org/10.1145/3472291.
Rožanec, J. M., Fortuna, B., Mladenić, D., 2022b. Knowledge Graph-Based Rich and
Confidentiality Preserving Explainable Artificial Intelligence (XAI). Information
Fusion 81 (December 2020), 91–102. https://doi.org/10.1016/j.inffus.2021.11.015.
Rožanec, J.M.; Fortuna, B.; Mladenić, D. 2022e. Reframing Demand Forecasting: A Two-
Fold Approach for Lumpy and Intermittent Demand. Sustainability 14 (15):9295.
(July 2022) https://doi.org/10.3390/su14159295
Rožanec, J. M., Novalija, I., Zajec, P., Kenda, K., Tavakoli, H., Suh, S., Veliou, E.,
Papamartzivanos, D., Giannetsos, T., Menesidou, S.A., 2022a. Human-Centric
Artificial Intelligence Architecture for Industry 5.0 Applications. http://arxiv.org/
abs/2203.10794.
Rožanec, J. M., Trajkova, E., Dam, P., Fortuna, B., Mladenić, D., 2022d. Streaming
Machine Learning and Online Active Learning for Automated Visual Inspection.
IFAC-PapersOnLine 55. Association for Computing Machinery. https://doi.org/
10.1016/j.ifacol.2022.04.206.
Rožanec, J. M., Trajkova, E., Novalija, I., Zajec, P., Kenda, K., Fortuna, B., Mladenić, D.,
2022c. Enriching Artificial Intelligence Explanations with Knowledge Fragments.
Future Internet 14 (5), 1–11. https://doi.org/10.3390/fi14050134.
Settles, B., 2010. Active Learning Literature Survey. Computer Sciences Technical Report
1648. https://doi.org/10.1016/j.matlet.2010.11.072.
Seung, H. S., Opper, M., Sompolinsky, H., 1992. Query by Committee, 287–294.
Shao, H., 2019. Query by Diverse Committee in Transfer Active Learning. Frontiers of
Computer Science 13 (2), 280–291. https://doi.org/10.1007/s11704-017-6117-6.
Shen, Y., Yun, H., Lipton, Z. C., Kronrod, Y., Anandkumar, A., 2017. Deep Active
Learning for Named Entity Recognition. Proceedings of the 2nd Workshop on
Representation Learning for NLP, Rep4NLP 2017 at the 55th Annual Meeting of the
Association for Computational Linguistics, ACL 2017, 252–.
Shi, X., Dou, Q., Xue, C., Qin, J., Chen, H., Heng, P-A., 2019. An Active Learning
Approach for Reducing Annotation Cost in Skin Lesion Analysis. Lecture Notes in
Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and
Lecture Notes in Bioinformatics) 11861 LNCS 628–636. https://doi.org/10.1007/978-
3-030-32692-0_72.
Shim, J., Kang, S., Cho, S., 2020. Active Learning of Convolutional Neural Network for
Cost-Effective Wafer Map Pattern Classification. IEEE Transactions on
Semiconductor Manufacturing 33 (2), 258–266. https://doi.org/10.1109/TSM.2020.
2974867.
Sinha, S., Ebrahimi, S., Darrell, T., 2019. Variational Adversarial Active Learning’
Proceedings of the IEEE International Conference on Computer Vision 2019-October:
5971–5980. https://doi.org/10.1109/ICCV.2019.00607.
Sugiyama, M., Kawanabe, M., 2013. Machine Learning in Non-Stationary Environments.
https://doi.org/10.7551/mitpress/ 9780262017091.001.0001.
116 Jože M. Rožanec, Blaž Fortuna and Dunja Mladenić

Taylor, A. T., Berrueta, T. A., Murphey, T. D., 2021. Active Learning in Robotics: A
Review of Control Principles. Mechatronics 77. https://doi.org/10.1016/j.mecha
tronics.2021.102576.
Tong, S., Chang, E., 2001. Support Vector Machine Active Learning for Image Retrieval.
Proceedings of the Ninth ACM International Conference on Multimedia -
MULTIMEDIA ’01 54 (C), 1–12.
Trajkova, E., Rožanec, J. M., Dam, P., Fortuna, B., Mladenić, D., 2021. Active Learning
for Automated Visual Inspection of Manufactured Products. Ljubljana ’21: Slovenian
KDD Conference on Data Mining and Data Warehouses, October, 2021, Ljubljana,
Slovenia. Association for Computing Machinery. http://arxiv.org/abs/2109.02469.
Wang, Z., Ye, J., 2015. Querying Discriminative and Representative Samples for Batch
Mode Active Learning. ACM Transactions on Knowledge Discovery from Data 9 (3),
17. https://doi.org/10.1145/2700408.
Weigl, E., Wolfgang, H., Lughofer, E., Radauer, T., Eitzinger, C., 2016. On Improving
Performance of Surface Inspection Systems by Online Active Learning and Flexible
Classifier Updates. Machine Vision and Applications 27 (1), 103–27. https://doi.org/
10.1007/s00138-015-0731-9.
Wu, D., 2019. Pool-Based Sequential Active Learning for Regression. IEEE Transactions
on Neural Networks and Learning Systems 30 (5), 1348–1359. https://doi.org/
10.1109/TNNLS.2018.2868649.
Wu, D., Lawhern, V. J., Gordon, S., Lance, B. J., Lin, C. T., 2016. Offline EEG-Based
Driver Drowsiness Estimation Using Enhanced Batch-Mode Active Learning
(EBMAL) for Regression. 2016 IEEE International Conference on Systems, Man, and
Cybernetics, SMC 2016 - Conference Proceedings, 730–736. https://doi.org/
10.1109/SMC.2016.7844328.
Wu, X., Chen, C., Zhong, M., Wang, J., Shi, J., 2021. COVID-AL: The Diagnosis of
COVID-19 with Deep Active Learning. Medical Image Analysis 68, 101913.
https://doi.org/10.1016/j.media.2020.101913.
Xie, B., Shen, G., Guo, C., Cui, Y., 2021. The Named Entity Recognition of Chinese
Cybersecurity Using an Active Learning Strategy. Wireless Communications and
Mobile Computing 2021. https://doi.org/10.1155/2021/6629591.
Xu, Z., Yu, K., Trespo, V., Xu, X., Wang, J., 2003. Representative Sampling for Text
Classification. Advances in Information Retrieval. ECIR 2003. Lecture Notes in
Computer Science 2633, 1–15.
Yang, L., Hanneke, S., Carbonell, J., 2013. A Theory of Transfer Learning with
Applications to Active Learning. Machine Learning 90 (2), 161–189. https://doi.org/
10.1007/s10994-012-5310-y.
Yu, H., Kim, S., 2010. Passive Sampling for Regression. Proceedings - IEEE International
Conference on Data Mining, ICDM, 1151–1156. https://doi.org/10.1109/ICDM.
2010.9.
Zajec, P., Rožanec, J.M., Trajkova, E., Novalija, I., Kenda, K., Fortuna, B., Mladenić, D.,
2021. Help Me Learn! Architecture and Strategies to Combine Recommendations and
Active Learning in Manufacturing. Information (Switzerland) 12 (11), 1–27.
https://doi.org/10.3390/info12110473.
Active Learning 117

Zhang, T., Oles, F., 2000. The Value of Unlabeled Data for Classification Problems.
Proceedings of the Seventeenth International Conference on Machine Learning
(ICML’00), 1191-1198.
Zhao, Y., Chen, D., Xie, H., Zhang, S., Gu, L., 2019. Mammographic Image Classification
System via Active Learning. Journal of Medical and Biological Engineering 39 (4),
569–82. https://doi.org/10.1007/s40846-018-0437-3.
Zhu, D., Li, Z., Wang, X., Gong, B., Yang, T., 2019. A Robust Zero-Sum Game Framework
for Pool-Based Active Learning. AISTATS 2019 - 22nd International Conference on
Artificial Intelligence and Statistics 89.
Zhu, J. J., Bento, j., 2017. Generative Adversarial Active Learning, 1–11. http://arxiv.org/
abs/1702.07956.
Zhu, X., Ghahramani, Z., Lafferty, J., 2003. Semi-Supervised Learning Using Gaussian
Fields and Harmonic Functions. Proceedings, Twentieth International Conference on
Machine Learning 2 (2001), 912–919.
Chapter 7

Prediction of General Anxiety Disorder

Using Machine Learning Techniques

Kevser Şahinbaş*
Management Information Systems, Istanbul Medipol University,
Istanbul, Turkey

Abstract

Today, the increase in mental health problems, the variable nature of

mental health and the lack of sufficient number of mental health
professionals have led to the search for machine learning that applied to
mental health problems extensively, and its use in the field of health is
considered as a new hope. Mental disorders are a health illness that
affects a person’s emotions, reasoning, and social interaction. Early
diagnosis and the application of the right treatment after the correct
diagnosis have always been the expectation of all humanity. As
technologies develop, machine learning has started to attract attention in
the field of medicine with the development of diagnostic methods. The
aim of this study is to conduct classification studies by using machine
learning methods in the diagnosis process of anxiety disorder diseases. A
publicly available dataset of 672 people’s Generalized Anxiety Disorder
7-item (GAD-7) responses during the COVID-19 period is used. This
study demonstrates that it is possible to classify mental health status with
0.97 accuracy rates with the Support Vector Machine algorithm, which
has a higher performance than other algorithms.

Keywords: general anxiety disorder, machine learning, data management,

classification

*
Corresponding Author’s Email: [email protected].

In: The Future of Data Mining

Editor: Cem Ufuk Baytar
ISBN: 979-8-88697-250-4
© 2022 Nova Science Publishers, Inc.
120 Kevser Şahinbaş

Introduction

Machine learning is increasingly becoming a part of digital medicine and

contributing to mental health research and practice. Machine learning can
improve diagnosis, monitoring, treatment, disease outcomes and rebalance
clinician workload. Anxiety disorders are one of the mental health problems.
In this respect, this study focuses on the prediction of anxiety disorders. The
aim of this study is to present a model that predicts anxiety disorder status in
the population of Bangladesh during the COVID-19 pandemic in terms of
GAD-7 score (Spitzer et al., 2006). This score is a symptom severity
evaluation and screening tool for the four most common anxiety disorders:
panic disorder, generalized anxiety disorder, social phobia, and post-traumatic
stress disorder, in primary care and mental health settings. On criteria like
work productivity and healthcare utilization, higher GAD-7 scores are linked
to disability and functional impairment. The severity of initial symptoms is
determined objectively and can be used to track symptom changes or the
effects of treatment over time.
COVID-19 has become a worry in developing nations like Bangladesh,
where the population is dense, and the health-care infrastructure is failing.
People were afraid due to a lack of protective and control capabilities.
Furthermore, they were losing money as a result of the emergency lockdown
(Bodrud-Doza, Shammi, Bahlman, Islam, & Rahman, 2020). Data suggests
that those confined in isolation or quarantine endure significant distress,
including anxiety, rage, bewilderment, and posttraumatic stress symptoms
(Brooks et al., 2020).
Early detection of mental health is a vital strategy for preventing and
responding to serious s problems and consequences caused by mental illness.
In this respect, machine learning-based technologies are making important
contributions to medicine and care management in various aspects. According
to many evidences, machine learning frameworks have the ability to speed up
the workflow of clinicians and other care providers. During the COVID-19
pandemic period, a research of the Bangladeshi population’s overall mental
health was conducted. The major goal of this study was to use survey questions
to predict anxiety disorder in people.
This chapter has following sections. The literature review follows the
introduction section and summarizes studies on mental health issues. The
background section provides information on the problem of mental health
prediction. The Materials and Methods section contains detailed analyses of
various ML algorithms such as Support Vector Machine (SVM), Decision
Prediction of General Anxiety Disorder Using Machine … 121

Tree (DT), Artificial Neural Network (ANN), K-Nearest Neighbor (KNN),

and Random Forest (RF). All algorithms’ analytical results are compared, and
the algorithm that obtains the best result is selected as the model. Finally, the
Results and Discussion section will go over the ML algorithms that have been
utilized to predict mental health concerns.

Literature Review

Xiong et al. (Xiong et al. 2021) proposed a model that apply Ensemble based
Bayesian Neural Network to predict three disorder by YODA dataset that
achieved 0.90 accuracy performance for Social Anxiety Disorder. In the
(Chekroud et al., 2016), the clinical remission from a 12-week regimen of
citalopram was predicted using a ML algorithm. The data was gathered from
1949 patients that suffer from level 1 depression with 25 variables to achieve
good performance. Their model achieved 64.6% accuracy rate by applying the
gradient boosting approach.
Ahmed et al. (Ahmed et al., 2020) presented a model for distinguishing
intensity level of the anxiety and depression at an early age by psychological
testing and some commonly used ML algorithms, namely CNN, KNearest
Neighbour, SVM and linear discriminant analysis. CNN obtained the best
classification performance with 96.8% for depression and 95% for anxiety.
Hilbert et al. (Hilbert et al., 2017) detected difficult cases from healthy cased
and GAD disorders from major depression by applying ML algorithms. They
used Binary SVM algorithm and achieved 90.10% accuracy.
Sau and Bhakta (Sau & Bhakta, 2017) proposed a model for detecting the
anxiety and depression in elderly patient by using ML algorithms such as
Naive Bayes, Random Forest, Bayesian network, Logistic regression and etc.
They achieved 89% accuracy rate with Random Forest algorithm by using 510
geriatric patients. Tat et al. designed a framework to predict mental health
problems in middle adolescence by using 474 predictors report and 7.638
twin’s data from the Child and Adolescent Twin Study in Sweden. The authors
applied ML algorithms, i.e., SVM, RF, XGBoost and Neural Network by
SMOTEBoost that is one of the imbalanced learning approaches and obtained
0.754 accuracy rate with RF algorithm.
Ćosić et al. (2020) addressed the issue of mental health disorders
prevention for healthcare professionals to predict a higher risk of chronic
mental health disorders during the COVID-19 pandemic. An objective
assessment of the intensity of exposure to stress and a self-report assessment
122 Kevser Şahinbaş

of health care workers was requested by obtaining data from clinical records
and hospital archives. The final stage included developing multimodal
stimulation paradigms to provide neurophysiological responses. At this stage,
the response to audio-visual stimuli was measured. Sensors such as fNIRS,
EEG, EKG, EMG, EDA were used for reaction measurements. Built-in eye
tracker, microphone and webcam were also available. It was emphasized that,
unlike statistical methods, feature selection and ML classification discovered
more complex nonlinear interactions. The study demonstrated formulability in
terms of supervised learning, both regression and classification tasks. The
neuro-physiological features retrieved at stage 4 can be found in a properly
formulated supervised learning task such as RF, SVM, ANN etc. It presented
a perspective with propositions such as its integrability with a model.

Background

Anxiety disorders are one of the most prevalent medical illnesses. Anxiety
disorders are common at both the population level and hospital cohorts and
are associated with a significant economic burden, poor quality of life, and a
variety of adverse outcomes (Kanwar et al., 2013). Anxiety disorders are a
common mental problem among teenagers (Rapee, Schniering, & Hudson,
2009). In addition to anxiety disorders due to general medical conditions,
generalized anxiety disorder, panic disorder, particular phobia, social anxiety
disorder and post-traumatic stress disorder are the five basic anxiety disorders,
as well as their linked conditions (Karamustafalıoğlu & Yumrukçal, 2011).
Anxiety is a type of affect with negative characteristics that distinguishes it
from other types of affect. Difficulty breathing, palpitations, rapid breathing,
trembling in the hands and feet, and excessive sweating are physiological
symptoms that can be characterized as psychological characteristics including
discomfort, excitement, emotion, and the fear of something bad happening
unexpectedly. Some definitions separate anxiety from fear by limiting the
source to the anticipation of an unknown danger.
Generalized anxiety disorder (GAD) is defined by the DSM-IV as a
condition characterized by intense and pervasive anxiety accompanied by a
variety of physical symptoms that causes significant impairment in social or
occupational functions or significant stress in the patient (Karamustafalıoğlu
& Yumrukçal, 2011).
Worry occurs almost every day for at least 6 months. It is about many
events or activities and is extreme. The individual has difficulty controlling
Prediction of General Anxiety Disorder Using Machine … 123

their anxiety. Accompanied by three or more of the following symptoms:

restlessness, excessive excitement, or anxiety, tiring easily, difficulty
concentrating or mind boggling, irritability, muscle tension, and sleep
disturbance (difficulty falling or staying asleep, or restless and restless sleep)
(Karamustafalıoğlu & Yumrukçal, 2011).
The high frequency of accompanying mental disorders makes it difficult
to predict the clinical course and prognosis. Panic disorder is seen in
approximately 25% of the patients, and major depressive disorder is also seen
in 50-80% of the patients. Coexistence of depression increases the risk of
suicide. It can continue throughout life (Sadock & Sadock, 2008).
It should be questioned whether the patient has been especially nervous
and worried that something bad will happen in the last months, what his/her
anxiety is about and whether he/she has been able to control his anxiety. Most
patients’ anxiety subsides when they have the opportunity to discuss their
difficulties with an attending physician. The clinician can identify external
situations that cause anxiety, change the environment with the help of the
patient or their family, and help them function effectively in their daily work
and relationships. Patients can thus reveal new rewards and pleasures that are
therapeutic for them (Karamustafalıoğlu & Yumrukçal, 2011).
GAD-7 seven-item measure that evaluates symptoms of generalized
anxiety disorder (GAD) showed reasonable specificity and sensitivity. This
scale includes seven questions with responses ranging from 0 (“Not at all”) to
3 (“Nearly every day”). The overall score was between 0 and 21. A total score
of 0–4 shows minimal anxiety, 5–9 shows mild anxiety, 10–14 shows
moderate anxiety, and 15–21 shows severe anxiety. Cronbach’s alpha was
0.84 for this scale. Anxiety positive was assigned to those who scored 10 in
this study score (Spitzer et al., 2006).
The diagnosis of generalized anxiety disorder should be investigated and
confirmed by other methods for those who have a total score of 10.

Materials and Methods

The dataset used is described, the SVM, Decision Tree, ANN, RF and KNN
algorithms are explained, and the experimental results are presented in detail
in this section.
124 Kevser Şahinbaş

Table 1. Generalized Anxiety Disorder scale (GAD-7)

Questions Score
1. In the last two weeks, I am feeling nervous, anxious, or on edge.
Not at all 0
Several days 1
More than half of the days 2
Nearly every day 3
2. In the last two weeks, I am not being able to stop or control worrying.
Not at all 0
Several days 1
More than half of the days 2
Nearly every day 3
3. In the last two weeks, I am worrying too much about different things.
Not at all 0
Several days 1
More than half of the days 2
Nearly every day 3
4.In the last two weeks, I feel trouble in relaxing.
Not at all 0
Several days 1
More than half of the days 2
Nearly every day 3
5. In the last two weeks, I am being so restless that it’s hard to sit still.
Not at all 0
Several days 1
More than half of the days 2
Nearly every day 3
6. In the last two weeks, I becoming easily annoyed or irritable.
Not at all 0
Several days 1
More than half of the days 2
Nearly every day 3
7. In the last two weeks, I am feeling afraid as if something awful might happen.
Not at all 0
Several days 1
More than half of the days 2
Nearly every day 3

SVM

SVM, developed by Vapnik, is a supervised learning method based on

statistical learning theory, named as Vapnik-Chervonenkis (VC) theory and
structural risk minimization, used for classification, regression, and pattern
recognition problems in data sets where patterns between variables are
Prediction of General Anxiety Disorder Using Machine … 125

unknown. SVM is distribution independent since it does not need any

combined distribution function information about the data (Cortes and Vapnik,
1995). Here denotes kernel functions, α Lagrange multipliers. The
inner products of the inputs are calculated with the help of kernel functions.
Lagrange multipliers represent the weights. In SVM, the output value for a
sample is equal to the sum of the dot products of the inputs and the independent
combinations of the Lagrange multipliers. The main purpose of SVM is to
separate the vectors belonging to different classes from each other and to
obtain the optimal separation hyperplane.

(1)

Decision Tree

In the algorithm created by Morgan and Songuist, sample data with known
classes is divided into small groups with simple decision-making steps. With
each division operation, similar data are grouped and classified by inductive
method (Safavian and Landgrebe, 1991). The basic step in decision trees is to
create decision nodes. While the decision nodes are being created, the best
attribute should be chosen as the node for the tree to branch in a balanced way
and for the classification process to be done correctly. For this, the expected
value in the whole system is calculated with the “Information Gain Theory”
revealed by Shannon and Weaver (Safavian and Landgrebe, 1991). The
information gain is calculated as in Equation 2:

(2)

𝑝𝑖: antecedent probability of class i

H: Entropy

Q: Sample space
𝑆𝑣: Sample space subset
126 Kevser Şahinbaş

ANN

Artificial Neural Networks (ANNs), which were first developed by Warren

McCulloch and W.A Pitts at the beginning of 1943 and inspired by the human
brain, make generalizations based on the data at hand and can make decisions
about the data they have never seen. Each nerve in ANNs has its own
information processing structure and is interconnected with other nerves
through weighted connections (Fyfe, 2000).
Inputs: ( ) brings the information received from the
environment to the nerve. Inputs can be added to the neural network from
previous nerves or from the outside world.
Weights: ( ) are suitable coefficients that decide the effect
of the inputs received by the artificial nerve on the nerve. Each entry has its
own weight.
Addition function: It adds the sums of the multiplication of each weight
in the nerve with the inputs it belongs to, with a threshold value, and sends it
to the activation function. In some cases, the aggregation function can be much
more complex, such as minimum, maximum, majority or several
normalization algorithms.
Activation function: The result of the addition is passed through the
activity function and passed to the output. The output of the event function 𝑦𝑖
is defined as in Equation (3) when adapted by the input vectors 𝑥𝑖.

If 1,
(3)
If 0

RF

Breiman introduces Random Forest (RF), a machine learning approach

originally designed for nonparametric multivariate classification (Catani et al.,
2013). Random Forest (RO) is a community classifier that collects the results
of a huge number of decision trees by majority vote (Kulkarni et al., 2016).
Random Forest classification, which is a decision tree classification
method, is an ensemble learning model in which more consistent results are
obtained by using more than one decision tree. This method, which is used in
both regression and classification problems, is a classifier that gives good
Prediction of General Anxiety Disorder Using Machine … 127

results without hyper-parameter estimation (Breiman, 2001). This method,

which creates more than one decision tree during the training, is input from
the results of these trees at the estimation stage. It enables the data to be
decided by majority vote (Breiman, 2001). This classifier gives successful
results because it solves the overfitting problem. The Random Forest
algorithm is created with ‘n’ decision trees. This algorithm performs
operations in two stages. In the first stage, a random forest model is created.
In the second stage, predictions are made from the created model with the help
of a classifier.

KNN

The k-nearest neighbor method calculates the class label of the test sample
with the labels of the test sample’s nearest neighbours (Soucy & Mineau,
2001). While classifying, the distances of each data in the data set to other data
are calculated. In order to determine this distance, k number of other records
is taken into consideration for a record. k recording distances are the closest
to the calculated point compared to other records. It is important to determine
the k value because the k value is too small to affect the model very much.
Although they are points of the same class, it causes some points to be placed
in separate classes or to create a separate class for those points. Likewise,
being too large causes it to be as if there is only one class, and dissimilar points
are classified together. For these reasons, it is seen that the number of k
influences the classification. 3, 5, 7, which refer to majority voting, are the
most used k values .
K-nearest neighbor algorithm steps;

Step 1: The new incoming sample is added to the class.

Step 2: Look at k neighbours
Step 3: The distance is calculated using the distance functions.
Step 4: The closest instance is assigned to that class.

Performance Metrics

True Positive (TP) indicates that the actual class and the predicted class have
the same value. TP value is found when we classify people without anxiety
disorder as non-anxious disorder. True Negative (TN) are correctly predicted
128 Kevser Şahinbaş

negative values. This indicates that the value of the actual class and the
predicted class are the same. When we classify those with anxiety disorders as
anxiety disorders, the TN value is found. The False Positive (FP) value appears
when the actual class and the predicted class overlap. FP value is found when
we classify a patient with anxiety disorder as non-anxious disorder. A False
Negative (FN) value appears when the actual class conflicts with the predicted
class. FN value is found when we classify a person who does not have an
anxiety disorder as an anxiety disorder. While it is desired to increase the true
positive and true negative areas, reducing the false positive and false negative
areas shows that the classification performance is good. The following metrics
can be calculated with the confusion matrix.

Table 2. Confusion Matrix

Predicted Class
Confusion Matrix
Non-Anxiety Disorder Anxiety Disorder
Non-Anxiety Disorder TN TP
Actual Class
Anxiety Disorder FN FP

The accuracy value, which is one of the performance measures of

classification models, shows the ratio of correctly predicted data to all
predicted values. It is obtained with the formula shown in 4 (Ögündür, 2019;
Witten et al., 2016).

(4)

Recall is the ratio of correctly classified positive data to total positive data.
From the data of the model created using this metric, the rate of finding
positive class labels is determined. Formula shown in 5 is calculated with
(Larose, D.T., 2014):

(5)

Unlike sensitivity, the ratio of positive samples classified correctly with

this criterion to the total positive predicted samples is measured by precision.
It is given at Eq. 6 (Karaağaoğlu et al., 2003):

(6)
Prediction of General Anxiety Disorder Using Machine … 129

F1 value, which is a harmonic mean of sensitivity and precision

performance measures, evaluates two different performance measures within
itself thanks to this feature. This metric gives a single benchmark (Eq. 7).
(Sokolova ve Japkowicz, 2006).

(7)

ROC is a measurement value that is frequently used in performance

measurements of machine learning algorithms. It shows how successful the
created model is in estimating. ROC is a probability curve and FPR on the x-
axis, y. On the axis, there are TPR values. Each point on the ROC curve
represents a sensitivity pair corresponding to a certain decision threshold
(Karaağaoğlu et al., 2003).

Data Description

The dataset used in this study is is publicly available1. The dataset represents
a one-of-a-kind look at mental health in Bangladesh’s population during the
early stages of the COVID-19 pandemic. The data were obtained from April
15 to May 10, 2020, when the Bangladeshi government implemented a curfew
and quarantine. A total of 672 (381 men, 291 women) individuals from the

Figure 1. GAD score count.

1
https://www.sciencedirect.com/science/article/pii/S2352340921006314?via%3Dihub.
130 Kevser Şahinbaş

Bangladeshi population aged 15-65 were included in the experiments and

analysis. In this study, the sociodemographic profiles of the individuals were
evaluated. The GAD-7 scale is used to assess anxiety. The questions in Table
1 were used to collect the dataset. Age, Gender, Marital Status, BMI,
Occupation, Education Level, Residence area, Economical Status, living
status, Smoking habits are features for sociodemographic profiles of the
individuals. Questions are shown in Table 1.
In Figure 1, GAD score class is presented.

Normalization Filter

Large differences between the data affect the learning accuracy of some
classification methods. The purpose of applying normalization is to facilitate
the comparison of data by eliminating the differences between mathematical
operations and data. In this study, the data were normalized using the Standard
Scaler in the Python programming language scikit-learn preprocessing library
in ANN, SVM, DT, KNN and RF algorithms.

Feature Importance

Figure 2 indicates the feature importance with RF algorithm. The results from
Figure 3 show that the question “In the last two weeks, I am feeling nervous,
anxious, or on edge” has the most impact on anxiety disorder prediction.
Marital status provides the least contribution.

Experiments and Results

In this section, Random Forest, Support Vector Machine, Artificial Neural

Networks, k-Nearest Neighbor algorithm, Random Forest classification
algorithms that are frequently preferred in classification models, are applied
to anxiety disorder data and the findings obtained as a result of the analysis
are indicated in detail.
The classifier model techniques given were tested using Python software
to determine the optimum model performance.
Prediction of General Anxiety Disorder Using Machine … 131

Figure 2. Feature importance.

ANN Results

In this study, multi-layer feedforward artificial neural network architecture

and Backpropagation method, which is the most widely used in prediction
problems, are used as the learning algorithm. The structure of the network is
defined as a multilayer perceptron consisting of 17 inputs, 2 output data and
hidden layers. 17 input sets are explained in the dataset section. The output set
is to have anxiety disorder and not have anxiety disorder. The neurons in the
hidden layer are between 50 and 100. Sigmoid for output layer and selu
function for hidden layers are used as activation function. The epoch number
is set at 5000. In ANN models, all other parameters are set as default values
in Python software.

Table 3. Performance metrics of the ANN

Model Accuracy Precision Recall f1-score

ANN 0.95 0.89 0.95 0.92
132 Kevser Şahinbaş

Table 3 presents performance metrics for ANN algorithm. Accuracy is

0.95, precision is 0.99, recall is 0.95 and F1-socre is 0.92.

Figure 3. History of ANN.

Figure 4. Accuracy history of ANN.

Prediction of General Anxiety Disorder Using Machine … 133

K-Nearest Neighbor Results

Table 4 provides KNN algorithm performance metrics. The accuracy,

precision, recall and F1 score values are 0.88, 0.91, 0.68 and 0.78 for KNN,
respectively. In figure 5, ROC Curve of KNN is presented.

Figure 5. ROC Curve of KNN.

Table 4. Performance Metrics of KNN Algorithm

Model Accuracy Precision Recall f1-score

KNN 0.88 0.91 0.68 0.78

Decision Tree Results

Table 5 presents performance metrics for Decision Tree algorithm. Accuracy

is 0.89, precision is 0.84, recall is 0.8 and F1-socre is 0.82.

Table 5. Performance Metrics of Decision Tree

Model Accuracy Precision Recall f1-score

Decision Tree 0.89 0.84 0.8 0.82
134 Kevser Şahinbaş

Figure 6. ROC Curve of Decision Tree.

Random Forest Results

The performance of the RF model is calculated, and the results are shown in
Table 6. Accuracy (0.95), precision (0.93), recall (0.9) and f1-score (0.91) are
obtained according to RF performance values.

Figure 7. ROC Curve of Random Forest.

Prediction of General Anxiety Disorder Using Machine … 135

Table 6. Performance Metrics of Random Forest

Model Accuracy Precision Recall f1-score

Random Forest 0.95 0.93 0.9 0.91

Support Vector Machine Results

During the analysing of SVM algorithm, the feature values that give the best
classification performance are adjusted (parameter tuning). The most ideal
kernel and attribute adjustment are set by the grid method as a result of the
findings. The Radial basis kernel is utilized as the kernel function in the SVM
model, and the gamma value () is set to 1 and C (cost parameter) is set to 1.

Table 7. Performance Metrics of SVM

Model Accuracy Precision Recall f1-score

SVM 0.97 1.0 0.91 0.95

The results in Table 7 demonstrate that the SVM algorithm has an obvious
advantage for anxiety disorder prediction. The accuracy, precision, recall and
f1-score values of the data analysed with SVM are 0.97, 1.0, 0.91 and 0.95,
respectively. The findings in Table 7 illustrate that the SVM algorithm
achieves the best performance in predicting anxiety disorder. In Figure 8, ROC
curve of SVM is illustrated.

Figure 8. ROC Curve of SVM.

136 Kevser Şahinbaş

Comparison of Performance Metrics for ANN, KNN, DT, RF, SVM

The findings in Figure 9 and Table 8 show that SVM algorithm achieves the
highest classification performance of predicting anxiety disorder with 0.97
accuracy performance.

Figure 9. Comparison of Performance Metrics.

Table 8. Performance Metrics of Algorithms

Model Accuracy Precision Recall f1-score

SVM 0.97 1.0 0.91 0.95
ANN 0.95 0.89 0.95 0.92
Random Forest 0.95 0.93 0.9 0.91
KNN 0.88 0.91 0.68 0.78
Decision Tree 0.89 0.84 0.8 0.82

Conclusion

In this study, different ML algorithms are applied to predict anxiety. This

research showed good screening properties for any anxiety disorder for the
GAD-7 questionnaire.
To unlock the full potential of machine learning, mental health scientists,
clinicians and patients must engage and take an active role in the clinical use,
communication and collaboration of machine learning to help transform
mental health practice and improve patient care.
Prediction of General Anxiety Disorder Using Machine … 137

The contribution of this study is to predict anxiety disorder by various

Machine Learning classification algorithms. Besides, feature importance by
RF is obtained. Among the algorithms which are ANN, KNN, DT, RF and
SVM used in this study, SVM algorithm achieved the best performance with
97% accuracy ratio. The question “In the last two weeks, I am feeling nervous,
anxious, or on edge” has the most influence on anxiety disorder prediction.
The fact that mental illnesses do not have a single cause, can be explained
by many reasons, and the overlap of many symptoms makes the diagnosis
process difficult for mental health professionals. However, with machine
learning, it is expected that the automatic assessment of the diagnosis of
thousands of people for psychiatric disease will be accessible by using
artificial intelligence algorithms with a large number of data and digital health
data records collected for years. Thus, a more accurate diagnosis process can
be made possible in a short time. In addition, with the collected data, the
prevalence of psychiatric illness in the population will be evaluated and
appropriate preventive and preventive studies can be carried out. Besides, it is
predicted that the collection of the mood follow-ups of the clients from data
collection environments such as digital diaries will increase the effectiveness
of the treatment process by providing continuous follow-up.
For the future work, it is planned to add additional mental health score
and adapt the questionnaire to Turkey.

References

Ahmed, A., Sultana, R., Ullas, M., Begom, M., Rahi, M., & Alam, M. A. (2020). A machine
learning approach to detect depression and anxiety using supervised learning. In 2020
IEEE Asia-Pacific Conference on Computer Science and Data Engineering (CSDE),
1-6.
Bland, R. C., Orn, H., & Newman, S. C. (1988). Lifetime prevalence of psychiatric
disorders in Edmonton. Acta Psychiatrica Scandinavica, 24-32.
Bodrud-Doza, M., Shammi, M., Bahlman, L., Islam, A. R., & Rahman, M. (2020).
Psychosocial and socio-economic crisis in Bangladesh due to COVID-19 pandemic: a
perception-based assessment. Frontiers in public health, 341.
Breiman, L., Random forests, 2001, Machine Learning, 45, 5–32.
Brooks, S. K., Webster, R. K., Smith, L. E., Woodland, L., Wessely, S., Greenberg, N., &
Rubin, G. J. (2020). The psychological impact of quarantine and how to reduce it:
rapid review of the evidenc. The lancet, 912-920.
Chekroud, A. M., Zotti, R. J., Shehzad, Z., Guerguieva, R., Johnson, M. K., & Corlett, P.
R. (2016). Cross-trial prediction of treatment outcome in depression: a machine
learning approach. The Lancet Psychiatry, 243-250.
138 Kevser Şahinbaş

Cortes, C. ve Vapnik, V., 1995, Support-vector networks, Machine Learning, 20, 273-97.
Fyfe, C., 2000, Artificial neural networks and information theory, http://index-of.co.uk/
InformationTheory/Neural%20Networks%20And%20Information%20Theory%20-
%20Colin%20Fyfe.pdf, [Access Date: 02 April 2022].
Hilbert, K., Lueken, U., Muehlhan, M., & Beesdo-Baum, K. (2017). Separating generalized
anxiety disorder from major depression using clinical, hormonal, and structural MRI
data: a multimodal machine learning study. Brain and behavior, e00633.
Kanwar, A., Malik, S., Prokop, L. J., Sim, L. A., & Murad, M. H. (2013). The association
between anxiety disorders and suicidal behaviors: A systematic review and meta‐
analysis. Depression and anxiety,, 917-929.
Karamustafalıoğlu, O., & Yumrukçal, H. (2011). Depresyon ve anksiyete bozuklukları.
Şişli Etfal Hastanesi Tıp Bülteni, 65-74.
Karaağaoğlu, E., Karakaya, J. ve Kılıçkap, M., 2016, Tanı Testlerinin Değerlendiril-
mesinde İstatistiksel Yöntemler/Statistical Methods in Evaluation of Diagnostic Tests,
Detay Publishing, Ankara.
Larose, D. T., 2014, Discovering knowledge in data: an introduction to data mining, 2nd
ed., John Wiley & Sons, Canada.
Rapee, R. M., Schniering, C. A., & Hudson, J. L. (2009). Anxiety disorders during
childhood and adolescence: Origins and treatment. Annual review of clinical
psychology, 311-341.
Sadock, B. J., & Sadock, V. A. (2008). Kaplan & Sadock’s concise textbook of clinical
psychiatry. Lippincott Williams & Wilkins.
Safavian, S. R. and D. Landgrebe, A survey of decision tree classifier methodology. IEEE
transactions on systems, man, and cybernetics, 1991, p. 660-674.
Sau, A., & Bhakta, I. (2017). Predicting anxiety and depression in elderly patients using
machine learning technology. Healthcare Technology Letters, 238-243.
Sokolova, M., Japkowicz, N. ve Szpakowicz, S., 2006, Beyond accuracy, f-score and roc:
a family of discriminant measures for performance evaluation, Australasian Joint
Conference on Artificial Intelligence, 1021.
Soucy, P., Mineau, G. W., 2001, A simple knn algorithm for text categorization, in: Data
Mining, ICDM, 647-648.
Spitzer, R. L., Kroenke, K., Williams, J. B., & Löwe, B. (2006). A brief measure for
assessing generalized anxiety disorder: the GAD-7. Archives of internal medicine,.
Archives of internal medicine, 1092-1097.
Witten, I. H., Frank, E., Hall, M. A. ve Pal, C. J., 2016, Data Mining: Practical machine
learning tools and techniques, 2nd ed., Morgan Kaufmann.
Xiong, H., Berkovsky, S., Romano, M., Sharan, R. V., Liu, S., Coiera, E., & McLellan, L.
F. (2021). Prediction of anxiety disorders using a feature ensemble based bayesian
neural network. Journal of Biomedical Informatics, 103921.
Editor’s Contact Information

Dr. Cem Ufuk Baytar, PhD

University Lecturer
Management Information Systems
Istanbul Topkapı University
Istanbul, Turkey
Email: [email protected]
Index

A C
active learning, v, viii, 95, 96, 97, 98, 100, churn customers, 51
101, 102, 103, 104, 105, 106, 107, 108, classification(s), 6, 31, 32, 33, 34, 35, 36,
109, 110, 111, 112, 113, 114, 115, 116, 37, 40, 41, 42, 43, 47, 48, 52, 58, 67, 69,
117 76, 85, 97, 98, 100, 102, 104, 107, 108,
algorithm(s), v, vii, 4, 6, 10, 12, 17, 29, 30, 109, 112, 113, 114, 115, 116, 117, 119,
31, 32, 33, 34, 35, 36, 37, 39, 40, 41, 42, 121, 122, 124, 125, 126, 127, 128, 130,
43, 45, 46, 47, 48, 49, 51, 52, 54, 55, 56, 135, 136, 137
57, 58, 59, 67, 69, 70, 71, 72, 73, 74, 75, crowdsourcing, v, viii, 77, 78, 82, 83, 92,
84, 86, 88, 96, 97, 99, 101, 102, 103, 93
104, 107, 109, 110, 113, 119, 120, 121, customer loyalty, 4, 29, 51, 52, 53, 54, 55,
123, 125,126, 127, 129, 130, 131, 132, 56, 57, 58, 76
133, 135, 136, 137, 138 customer relations, 30, 51, 54, 55, 56, 57,
analysis, v, vii, 2, 3, 4, 5, 6, 8, 9, 10, 11, 58, 74, 75, 76
13, 15, 16, 17, 18, 19, 21, 22, 23, 24, 25, customer(s), v, vii, 4, 5, 10, 22, 29, 30, 31,
26, 30, 33, 34, 41, 46, 48, 51, 52, 53, 54, 33, 34, 39, 46, 48, 51, 52, 53, 54, 55, 56,
55, 56, 57, 58, 59, 66, 74, 75, 76, 82, 84, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 69,
85, 86, 89, 106, 112, 115, 116, 121, 130, 73, 74, 75, 76, 105
138 cybersecurity, 98, 109, 110, 116
anxiety disorder, 119, 120, 121, 122, 123,
124, 127, 128, 130, 131, 135, 136, 137, D
138
application(s), v, vii, 2, 9, 10, 11, 12, 14, data analysis, 3, 16, 49, 51, 52, 54, 55, 57,
15, 16, 21, 24, 25, 27, 29, 30, 31, 33, 36, 58, 66, 75, 84
41, 48, 49, 74, 75, 76, 85, 86, 95, 97, data analytics, v, 1, 2, 3, 4, 11, 12, 58, 92
109, 111, 112, 115, 116, 119 data management, 2, 84, 119
artificial neural network(s) (ANNs), 6, 52, data mining, v, vii, 2, 3, 5, 10, 11, 12, 13,
56, 58, 67, 71, 72, 73, 74, 75, 121, 126, 15, 17, 18, 19, 21, 22, 24, 25, 26, 29, 30,
130, 131, 138 31, 32, 33, 35, 37, 39, 47, 48, 49, 51, 52,
54, 56, 57, 58, 59, 67, 73, 75, 76, 77, 78,
82, 84, 85, 86, 92, 111, 116, 138
B
data preparation, 58, 66, 67
banking, v, vii, 51, 53, 57, 58, 74, 76 database, vii, 3, 5, 14, 15, 16, 18, 24, 25,
bioinformatics, 13, 14, 20, 24, 25, 26, 47, 28, 57, 75, 84
112, 114, 115
142 Index

decision tree, vii, 6, 12, 30, 31, 32, 34, 35, prioritization, v, vii, 13, 14, 15, 16, 20, 21,
36, 39, 42, 43, 44, 45, 46, 47, 48, 52, 56, 22, 24
58, 67, 68, 69, 70, 73, 74, 85, 107, 121, promising, v, vii, 13, 14, 15, 21, 22
123, 125, 126, 133, 134, 136, 138 prosumer, v, viii, 77, 78, 79, 82, 84, 89, 93

G R
general anxiety disorder, vi, viii, 119 random forest (RF), vii, 30, 32, 33, 35, 36,
42, 43, 44, 45, 46, 47, 48, 49, 52, 58, 67,
H 69, 70, 74, 75, 76, 121, 122, 123, 126,
130, 134, 135, 136, 137
healthcare, 10, 95, 106, 110, 120, 121, 138 robotic(s), 95, 104, 106, 110, 113, 116
human resources, vii, 1, 2, 7, 9, 11
human resources metrics, 2, 10
S
I software, vii, 3, 14, 15, 16, 17, 18, 19, 27,
32, 34, 35, 37, 39, 48, 52, 130, 131
in silico, 14, 17, 18, 22, 23, 24, 25, 26, 27 strategy(ies), vii, 1, 8, 10, 19, 21, 29, 33,
industry, v, viii, 1, 31, 48, 51, 52, 54, 58, 34, 46, 54, 55, 56, 58, 75, 76, 79, 92, 97,
75, 84, 93, 104, 105, 115 98, 99, 101, 103, 104, 105, 108, 109,
110, 111, 112, 113, 116, 120
K support vector machine (SVM), 35, 100,
k-means, 6, 10, 30, 31, 36, 103, 106 101, 108, 111, 114, 116, 119, 120, 121,
k-nearest neighbor (KNN), 6, 121, 123, 122, 123, 124, 130, 135, 136, 137
127, 130, 133, 136, 137
T
L technique(s), vi, viii, 2, 3, 4, 5, 10, 11, 12,
learning from demonstration, 95 13, 14, 20, 22, 30, 32, 41, 48, 49, 52, 76,
78, 84, 85, 86, 92, 95, 96, 99, 100, 103,
104, 119, 130, 138
M
tool(s), ii, v, vii, 4, 5, 7, 13, 14, 15, 16, 18,
machine learning, vi, viii, 4, 10, 11, 34, 35, 21, 22, 24, 25, 29, 33, 35, 48, 57, 58, 69,
47, 49, 67, 69, 76, 95, 96, 97, 99, 105, 77, 78, 105, 120, 138
109, 110, 111, 112, 114, 115, 116, 117, toxicogenomic(s), v, vii, 13, 14, 15, 18, 19,
119, 120, 126, 129, 136, 137, 138 20, 21, 22, 23, 24, 25, 26, 27, 28
manufacturing, 33, 95, 98, 104, 105, 110, true negative (TN), 38, 43, 68, 70, 73, 127
111, 112, 113, 114, 115, 116 true positive (TP), 37, 42, 43, 68, 70, 73,
microgrid(s), v, viii, 77, 78, 79, 80, 81, 82, 127, 128
85, 86, 88, 89, 90, 91, 92, 93
X
P
x-means, 30, 31
prediction, vi, viii, 6, 20, 22, 26, 32, 48, 49,
56, 76, 99, 105, 112, 114, 119, 120, 130,
131, 135, 137, 138