Auto-weighted multi-view co-clustering with bipartite graphs

Abstract

Co-clustering aims to explore coherent patterns by simultaneously clustering samples and features of data. Several co-clustering methods have been proposed in the past decades. However, in real-world applications, datasets are often with multiple modalities or composed of multiple representations (i.e., views), which provide different yet complementary information. Hence, it is essential to develop multi-view co-clustering models to solve the multi-view application problems. In this paper, a novel multi-view co-clustering method based on bipartite graphs is proposed. To make use of the duality between samples and features of multi-view data, a bipartite graph for each view is constructed such that the co-occurring structure of data can be extracted. The key point of utilizing the bipartite graphs to deal with the multi-view co-clustering task is to reasonably integrate these bipartite graphs and obtain an optimal consensus one. As for this point, the proposed method can learn an optimal weight for each bipartite graph automatically without introducing an additive parameter as previous methods do. Furthermore, an efficient algorithm is proposed to optimize this model with theoretically guaranteed convergence. Extensive experimental results on both toy data and several benchmark datasets have demonstrated the effectiveness of the proposed model.