Similarity-based spectral clustering ensemble selection

Abstract

Traditional clustering ensemble methods combine all the obtained clustering results at hand. However, we can often achieve a better clustering solution if only a part of all available clustering results is combined. In this paper, SELective Spectral Clustering Ensemble (SELSCE), a novel clustering ensemble method, is proposed. The component clusterings of an ensemble system are generated by Spectral Clustering (SC) capable of engendering diverse committees. The random scaling parameter, Nystrom approximation and random initialization of k-means are used to perturb SC for producing the components of an ensemble system. A measure integrating diversity and quality is proposed to evaluate the quality of all the obtained results. Furthermore, a novel selection strategy based on the nearest neighbor rule is introduced to choose from them a part of the promising to build an ensemble committee. The experimental results on UCI datasets demonstrate that the proposed algorithm outperforms the traditional ones in data clustering.