Does there exist a flavor of spectral clustering (graph construction, type of laplacian, et cetera) that results in a clustering (on arbitrary data sets) that is similar to the k-means clustering "in spirit", i.e. finds a partition of the data into isotropic high density regions? (Of course, I don't expect exact equivalence.)

Whenever I've seen spectral clustering, it has seemed to go more after manifolds (i.e. non-isotropic clusters), but that may be because commonly spectral clustering is presented using k-NN type graphs. Of course, a k-NN (or epsilon-neighborhood) graph contains no large-scale information on the relative positions of points, hence it is in no way "encouraged" towards isotropy.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.