Fast Dynamic Multi-view Clustering with semantic-consistency inheritance

Multi-view clustering has attracted much attention in recent years. However, most current methods are designed to handle fixed multi-view data, overlooking practical scenarios where data views are collected sequentially. Although some researches have attempted to handle the streaming views, certain limitations are often encountered. (1) Some methods focus on exploring pairwise similarity relations between data for clustering, resulting in high computational costs. (2) A similar latent representation is typically assumed for all distinct views, neglecting the detrimental impact of non-semantic information present in each view. To address these issues, a novel Fast Dynamic Multi-view Clustering (FDMVC) method with semantic-consistency inheritance is proposed. In specific, FDMVC efficiently learns a latent representation by matrix factorization for each view, and further disentangles it into a low-rank semantic-consistent part and a view-specific non-semantic part. When new views come, only the semantic-consistent representation is reused for knowledge transfer. This approach mitigates the negative impact of non-semantic information in previous views. The proposed method can be solved efficiently with linear time and space complexity. Experiments demonstrate the effectiveness and efficiency of the proposed approach compared with state-of-the-art methods.