Finding the number of clusters on the basis of eigenvectors

Abstract

Finding the number of clusters is a challenging task. We suggest a new method for an assessment of a group number. Our solution uses only simple properties of signless Laplacian eigenvectors. The novel method has been incorporated to our previous spectral algorithm. The performance of the modified version is competitive to existing solutions. We empirically evaluate the proposed approach using standard test sets and show that it is able to find correct partitioning even for weakly separated groups of varying densities. © 2013 Springer-Verlag Berlin Heidelberg.