Joint learning of data recovering and graph contrastive denoising for incomplete multi-view clustering

Abstract

Incomplete multi-view clustering is pivotal in machine learning because complex systems are inherently difficult to be fully observed and therefore pose a great challenge to revealing the mechanisms and structure of underlying systems. Current methods are criticized for disregarding features with missing views or inadequately exploring the local structure within incomplete views. To solve these limitations, we propose a novel Unified Graph Contrastive Learning Framework for incomplete multi-view clustering (called UGCF), which jointly learns data restoration, graph contrastive denoising, and clustering. Specifically, UGCF first restores the missing values by exploiting the conserved relations in each view, and the local structure within data is preserved. Second, UGCF removes the heterogeneity of multi-view data by learning an affinity structure of objects for each view and constructs a unified graph for multi-view data by manipulating the topological structure of affinity graphs. To enhance the quality of the features of vertices, graph contrastive learning is executed on the unified graphs by selecting positive and negative samples, which considerably improves the discriminative of features, thereby eliminating noise in data. Finally, UGCF integrates data restoration, graph contrastive denoising, and clustering into an overall objective. Experiments demonstrate the superiority of UGCF over state-of-the-art baselines in incomplete multi-view clustering with various metrics.