A Spectral Clustering Algorithm Based on Eigenvector Localization

This paper introduces the SpecLoc algorithm that performs clustering without pre-assigning the number of clusters. This is achieved by the use of a special property of matrix eigenvectors, called weak localization. The signless Laplacian matrix is created on the basis of a mutual neighbor graph. A new measure, introduced in this work, allows for selection of weakly localized eigenvectors. Experiments confirm good performance of the proposed algorithm for weakly separated groups of real datasets, including cancer gene expression matrices.