Model Selection and Stability in Spectral Clustering
Zeev Volkovich, Renata Avros · 2012
An open problem in spectral clustering concerning of finding automatically the number of clusters is studied. We generalize the method for the scale parameter selecting offered in the Ng-Jordan-Weiss (NJW) algorithm and reveal a connection with the distance learning methodology. Values of the scaling parameter estimated via clustering of samples drawn are considered as a cluster stability attitude such that the clusters quantity corresponding to the most concentrated distribution is accepted as true number of clusters. Numerical experiments provided demonstrate high potential ability of the offered method.