Non-Negative Half-Space Clustering with Sparseness Constraints

Abstract

This paper describes a novel clustering approach by revealing the non-negative half-space clustering with sparseness constraints (NHCS). Sparseness can make only few components of whole samples to be ‘active’. Especially, this method is more part-based compared to other matrix factorization methods, which is sensitive to the scale of the data. After obtaining the part-based structure, the samples can be grouped by spectral cutting techniques. It shows that our method has more robust with the increasing of the number of clusters. Both theoretical and experimental results show that NHCS performs better than other competitive algorithms on the two database CBCL and Reuters-21578.