Direct Alignment Maximization for Clustering Ensemble

Liang Du, Kang Yang, Yunhui Liang, Xiaolin Lv

Abstract: Given multiple clustering algorithms with different candidate hyper-parameter configurations, we will generate multiple weak partition results for the same data set. By integrating a group of such base clustering results, clustering ensemble methods try to find the final consensus clustering result where the stability and even the performance of base clustering results could be further improved. Although the exact one-to-one mapping across clusters from different base clustering may not exist, we still have chance to directly find such weak correspondence between the final consensus clusters and the input base clusters. In this paper, we propose a novel method which maps all the clusters from base clustering to new permuted clusters and finds the consensus clustering by maximizing the alignment from permuted bases. It has been verified that the proposed method outperforms several state-of-the-art baseline methods on a variety of data sets.