auto_spectral_clustering

Spectral clustering is one of the most popular modern clustering algorithms. Typically spectral clustering requires number of clusters manually.

auto_spectral_clustering predicts number of clusters based on the eigengap(often referred to as spectral gap).

References

A Tutorial on Spectral Clustering, 2007, Luxburg, Ulrike[1] http://www.kyb.mpg.de/fileadmin/user_upload/files/publications/attachments/Luxburg07_tutorial_4488%5b0%5d.pdf

License

New BSD License.

Notes

The eigengap heuristic usually works well if the data contains very well pronounced clusters, but in ambiguous cases it also returns ambiguous results.[1]

Dependencies

auto_spectral_clustering is tested under Python 3.3.5.

It requires NumPy, SciPy, scikit-learn. If you test it, it also requires matplotlib.

Usage

from autosp import predict_k
from sklearn.cluster import SpectralClustering

k = predict_k(affinity_matrix)
sc = SpectralClustering(n_clusters=k,
                        affinity="precomputed",
                        assign_labels="kmeans").fit(affinity_matrix)

labels_pred = sc.labels_

Examples

You can change number_of_blobs(artificial datasets) and test it!!

test.py

if __name__ == "__main__":

    # Generate artificial datasets.
    number_of_blobs = 5  # You can change this!!
    data, labels_true = make_blobs(n_samples=number_of_blobs * 10,
                                   centers=number_of_blobs)