Diversity-induced consensus and structured graph learning for multi-view clustering

Graph-based multi-view clustering has recently attracted extensive attention due to its capacity of exploring the nonlinear structure of data points from multiple views. However, most existing methods focus on discovering the consistency of multiple views, while ignoring the inconsistencies that may be caused by noise, outliers or view-specific attributes. To this end, this paper proposes a Diversity-induced Consensus and Structured Graph Learning model for multi-view clustering (DCSGL), which simultaneously formulates the multi-view consistency and the multi-view diversity into a unified framework to guide the consensus and structured graph learning. Specifically, DCSGL decomposes each view initial graph into two latent factors, i.e., consistent factor and inconsistent factor, and then adopts the Hilbert Schmidt Independence Criterion (HSIC) as a diversity penalty term to force inconsistent factors to be sparse across views, thereby inducing consistent factors to be cleaner than initial graphs. Furthermore, the consistent factors of different views are fused into a consensus graph with an explicit connectivity constraint in a self-weighted manner, leading to the components in the consensus graph indicate clusters directly. Experimental results on seven benchmark datasets demonstrate the effectiveness of our proposed method, comparing to the state-of-the-art methods for multi-view clustering.