Multi-View Clustering With Learned Bipartite Graph

Abstract

In this paper, we propose a late fusion multi-view clustering via a learned bipartite graph (MVC-LBG). Firstly, we obtain the base partition from each single view. Then, a consensus bipartite graph is constructed by these base partitions for final clustering results. Inspired by anchors used in spectral clustering on a bipartite graph, we introduce representative points that are updated during the iterating of MVC-LBG. Compared with early fusion multi-view clustering algorithms, MVC-LBG not only dramatically improves the clustering accuracy but also has lower computational complexity. By adopting the base partitions in a new way, the proposed algorithm achieves better performance than existing late fusion multi-view clustering algorithms. Extensive experiments on six data sets demonstrate the effects and efficiency of our algorithm. Finally, we extend MVC-LBG to large-scale datasets.