Feature Selection in Spectral Clustering

Abstract

Spectral clustering is a powerful technique in clustering specially when the structure of <br><br>data is not linear and classical clustering methods lead to fail. In this paper, we propose a <br><br>spectral clustering algorithm with a feature selection schema based on extracted features of <br><br>Kernel PCA. In the proposed algorithm, selecting appropriate vectors is dependent upon <br><br>entropy of clusters on these vectors and weighting method is influenced by sum of the <br><br>existence gap between clusters and entropy of the vectors. Tuning the parameters has a great <br><br>effect on the results of spectral clustering techniques. In the ideal case, comparing our <br><br>method with NJW and Kernel K-Means indicate the successful of the proposed algorithm.