Tensorized graph-guided view recovery for incomplete multi-view clustering

Abstract

Multi-view clustering (MVC) methods have demonstrated remarkable success when all samples are available across multiple views by leveraging consistency and complementary information. However, real-world multi-view data often suffers from incompleteness, where some samples are missing in one or more views. This incompleteness makes MVC challenging, as it becomes difficult to uncover consistency and complementary relationships among the view data. As a result, Incomplete Multi-View Clustering (IMVC) has emerged to address the limitations posed by missing data. An intuitive approach to tackle this issue is view recovery-effectively leveraging consistency information from multiple views to impute missing data. However, the quality of view recovery heavily depends on the learned consistency information, making it crucial to learn high-quality consistency representations. To address this challenge, we propose a novel approach called Tensorized Graph-Guided View Recovery (TGGVR), which integrates view recovery and tensorized graph learning within a unified framework. The tensorized graph learning estimate a similarity graph for each view by exploiting consistency and complementary information through tensorized learning. In addition, high-quality neighborhood structures are exploited to obtain a more accurate consensus graph. This high-quality consensus graph then guides the more accurate recovery of missing data, establishing a cyclical procedure in which tensorized graph learning and data imputation mutually reinforce each other. Experimental results demonstrate that our proposed method outperforms several state-of-the-art approaches in tackling the challenging task of IMVC. Notably, our method significantly outperforms representative competing methods by more than 5% and 10% on the BBC and Caltech datasets, respectively.