I thought I understood Spectral Clustering well enough till I came across these two paragraphs:
Graph Laplacians are interesting linear operators attached to graphs. They are the discrete analogues of the Laplace-Beltrami operators that appear in the Hodge theory of Riemannian manifolds, whose null spaces provide particularly nice representative forms for de Rham cohomology. In particular, their eigenfunctions produce functions on the vertex set of the graph. They can be used, for example, to produce cluster decompositions of data sets when the graph is the 1-skeleton of a Vietoris-Rips complex. We ﬁnd that these eigenfunctions (again applied to the 1-skeleton of the Vietoris-Rips complex of a point cloud) also can produce useful ﬁlters in the Mapper analysis of data sets
– From Prof. Gunnar Carlsson’s survey Topology and Data. (More on this survey as a manifesto for “Topology and Data” in a later post). That aside, I do like how the image on the wiki entry for Vietoris-Rips complex looks like:
A little less intimidating ( now this almost borders on “ofcourse that’s how it is”. I am interested in the same reaction for the paragraph above some months later):
A related application [of the graph laplacian] is “Spectral Clustering”, which is based on the observation that nodal domains of the first eigenvectors of the graph laplacian can be used as indicators for suitably size-balanced minimum cuts.
– From Laplacian Eigenvectors of Graphs linked in the previous post. While this isn’t really as compressed as the lines above, they made me think since I did not know about Courant’s Nodal domain theorem. Like I did in the previous blog post, I would highly recommend this (about 120 page) book. It soon covers the Nodal Domain theorem and things make sense (even in context of links between PCA and k-means and Non-Negative Matrix Factorization and Spectral Clustering, at least in an abstract sense).