Feature Selection Criteria for Clustering Based on Support Vectors

Existing feature selection for clustering task usually relies on some statistical methods or some clustering quality evaluation, both of which need an assumption of data distribution. That heavily affects the generalization ability of feature selection in applications. This paper proposes feature selection criteria derived from support vector technique to address this problem. Firstly the least power support vector clustering (LPSVC) is presented. Based on products of LPSVC, an institutive criterion and a derivative criterion are proposed. These two LPSVC-based criteria keep data distribution shape as much as possible after features are removed. And they are equipped with selection algorithm to discard unimportant features. Empirical evidence suggests the proposed criteria are of fine quality.