Spectral Clustering Ensemble Based on Synthetic Similarity

Abstract

In this paper, a spectral clustering ensemble algorithm based on synthetic similarity (SCEBSS) is proposed to improve the performance of clustering. Multiple methods of vector similarity measurement are adopted to produce diverse similarity matrices of objects. Every similarity matrix is given a weight and then added as a synthetic similarity matrix. A spectral clustering algorithm is employed on the synthetic similarity matrix, and then a particle swarm optimization using normalized mutual information (NMI) as evaluation function is adopted to optimize the weights of similarity matrices to obtain the best clusters. Comparisons with other related clustering schemes demonstrate the better performance of SCEBSS in clustering data tasks and robustness to noise.