Clustering ensemble based on multiscale cluster reliability

Abstract

Clustering ensemble can effectively improve the accuracy and robustness of clustering. The quality and diversity of base clustering are crucial to the clustering ensemble effect. However, traditional clustering ensemble algorithms usually treat a base clustering result as a whole, ignoring the quality differences within the same base clustering result. Clusters of different quality have completely different contributions to clustering ensemble results, and ignoring the quality differences of clusters can seriously affect the effectiveness of clustering ensemble. To solve this problem, this paper proposes a clustering ensemble method based on multiscale cluster reliability. In this method, multiscale global structural information is first mined by performing random walks on the cluster similarity graph to obtain multiscale correlation between clusters. Then, the multiscale cluster reliability is measured by using the multiscale correlation of clusters, combined with the information entropy, and the clusters are weighted accordingly. Lastly, the clustering results are obtained by graph segmentation. Experimental results on numerous real datasets show that compared with 10 typical and state-of-the-art clustering ensemble methods, the clustering ensemble method based on multiscale cluster reliability not only performs better in clustering effect but also has higher stability.