Localization Graph Generation Based Matrix Factorization for Incomplete Multi-View Clustering

Incomplete multi-view clustering (IMC) has received increasing attention since missing observations of views are common in real-world applications. Existing approaches often learn similarity graphs from raw data, but the incompleteness of raw data limits their ability to capture complete similarity. Moreover, the local structure of the complete similarity graph is rarely considered. To alleviate these problems, in this paper, we propose a novel graph-based IMC method by designing a localization graph generation strategy on the weighted multi-view matrix factorization. Specifically, it first derives a consensus representation with balanced sample importance by the sample importance-weighted matrix factorization. Then, instead of learning graph from raw data, it develops a localization graph generation strategy to generate a complete consensus graph from the consensus representation. Especially, the dot-product weighting and the 0-1 weighting are learned adaptively, such that the local structure of the consensus graph is preserved well. Besides, the representation learning and localization graph generation are integrated into one joint optimization problem, where they boost each other. To optimize the proposed model, we provide an effective algorithm with a lemma and a theorem about its convergence. Finally, experimental results on several real-world datasets show the effectiveness and advancement of the proposed method, compared to the state-of-the-art methods.