Graph-filtering and high-order bipartite graph based multiview graph clustering

Abstract

Multiview clustering, which partitions data into different groups, has attracted wide attention. With increasing data, bipartite graph-based multiview clustering has become an important topic since it can achieve efficient clustering by establishing relationship between data points and anchor points instead of all samples. Current most methods learn graph structure using one-order bipartite graph from original multi-view feature. However, original data inevitably contain noise feature and its structure is complex. To address this issue, a novel graph filtering and high-order bipartite graph-based multiview graph clustering method is presented, which consider the influence of noise feature and complex graph structure relationship. Concretely, we first employ graph filtering to original data feature space, then adopt a two-step random walk approach to establish the bipartite graph structural relationships of each view. At last, based on constructed high-order bipartite graph, a self-weight bipartite graph-based multiview graph fusion framework is proposed, which reduces annoying weight parameter selection and obtains a joint bipartite graph. Experimental results on several benchmark datasets demonstrate that this method achieves better clustering performance than state-of-the-art multiview clustering methods.