Attribute Subset Kernels Reflecting Discernibility of Clusters

Abstract

In this paper, we study kernel functions for clustering with nominal data sets, which are consisting of objects described by nominal attributes. For such data sets, it is important that clusters have simple representations of the attributes, as well as clusters of similar objects are obtained. To obtain such clusters, we propose a kernel function defined through Boolean formulas which reflect discernibility of clusters. We apply the proposed kernel function to clustering under cannot-link constraints.