Consensus Multi-view subspace clustering based on Graph Filtering

Abstract

Multi-view subspace clustering has attracted an increasing amount of attention because it can capture information from multiple views as well as avoid the curse of dimensionality. The existing methods cannot simultaneously achieve both the effective reduction of negative impact of noise and the high-quality consensus representation within a unified framework. To handle this issue, this paper proposes a novel Consensus Multi-view subspace clustering based on Graph Filtering, named CMGF. First, CMGF learns a latent data space by using view-specific k-order filters to reduce noise and redundant information. Then, CMGF obtains the spectral embedding matrix of each view by imposing graph regularization constraint. Ultimately, to generate a consensus representation, we integrate the spectral embedding matrix of each view by using an adaptively weighted scheme. Experimental results on ten real-word datasets show that the proposed method outperforms state-of-the-art baselines significantly. Experiments also demonstrate that the graph filtering employed in CMGF enhances the smoothness of the data and improves the distinctiveness of the cluster structure.