Ensemble clustering by block diagonal representation

Ensemble clustering integrates all basic clustering results to produce a better clustering result. Existing ensemble clustering methods typically rely on a co-association matrix (CA), which measures the number of occurrences two samples are grouped into the same cluster in different base clusters. However, ensemble clustering performance degrades when the generated CA matrix is of low quality. In this paper, we improve the quality of CA matrix by block diagonal regularization to obtain better clustering performance, and propose ensemble clustering by block diagonal representation (BEC). Specifically, for a given CA matrix, We decompose the CA matrix into two parts, one is the error matrix which is sparse, and the other is the new CA matrix, which is required to be a block diagonal matrix. Since BEC uses the block diagonal structure prior, we obtain a new CA matrix with high quality. In addition, a rank constraint is imposed to the Laplacian matrix of the new CA matrix, such that the connected components in the CA matrix are exactly equal to the cluster number. Thus the final clustering result can be directly obtained from the new CA matrix. The experimental results of different clustering algorithms on 9 benchmark datasets show the effectiveness of the proposed model in ensemble clustering.