Clustering by Learning the Non-Negative Half-Space

Abstract

This paper proposes a novel clustering algorithm which is called Non-negative Half-space Clustering (NHC), by revealing the nonnegative half-space structure of samples. The half-space is the union of some nearly independent half-spaces, and each class of samples is dominated by this half-space. Since the subspace independent assumption is not imposed on the samples, NHC is robust for the increasing of number of classes compared with other subspace clustering methods such as Sparse Space Clustering. After obtaining a half-space structure, the adjacency graph is almost block-wise, and can be well grouped by some cutting techniques. In the experiment section, we implement NHC and other competitive algorithms on two database CBCL and Reuters-21578. The result shows that NHC performs better on the two database, and more robust than SSC.