Save the spectral embeddings used for clustering in the SpectralClustering class and make them accessible through an attribute, e.g. maps_, to make easier post-processing on the clusters.

@matteo-bastico it would help us decide if you could explain how you would use this.

As a workaround, you may recompute the mapping after fitting SpectralClustering:

from sklearn.cluster import SpectralClustering
from sklearn.manifold import spectral_embedding
import numpy as np

X = np.array([[1, 1], [2, 1], [1, 0],
              [4, 7], [3, 5], [3, 6]])
clustering = SpectralClustering(n_clusters=2,
        assign_labels='discretize',
        random_state=0).fit(X)

maps = spectral_embedding(
    clustering.affinity_matrix_,
    n_components=clustering.n_clusters,
    eigen_solver=clustering.eigen_solver,
    random_state=0,
    eigen_tol=clustering.eigen_tol,
    drop_first=False,
)

Save the spectral embeddings used for clustering in the SpectralClustering class and make them accessible through an attribute, e.g. maps_, to make easier post-processing on the clusters.

@matteo-bastico it would help us decide if you could explain how you would use this.

In my case, I want to compute the medoids of the clusters using the distances in the spectral embedding space instead of the original Euclidean space. But there are others applications in which this feature can be useful.

As a workaround, you may recompute the mapping after fitting SpectralClustering:

from sklearn.cluster import SpectralClustering
from sklearn.manifold import spectral_embedding
import numpy as np

X = np.array([[1, 1], [2, 1], [1, 0],
              [4, 7], [3, 5], [3, 6]])
clustering = SpectralClustering(n_clusters=2,
        assign_labels='discretize',
        random_state=0).fit(X)

maps = spectral_embedding(
    clustering.affinity_matrix_,
    n_components=clustering.n_clusters,
    eigen_solver=clustering.eigen_solver,
    random_state=0,
    eigen_tol=clustering.eigen_tol,
    drop_first=False,
)

Thank you, as a workaround it works but the spectral embeddings are computed twice and for large matrices it is time consuming.