Graph Embedding-Based Deep Multi-view Clustering

Abstract

The graph embedding techniques adeptly leverage the neighbor prior information within the original high-dimensional sample space, imposing constraints on the low-dimensional representational space to ensure that the latent features accurately reflect the connectivity strength of the original samples. In the domain of single-view clustering, graph embedding methods have been widely applied to some traditional clustering algorithms to obtain robust clustering performance. However, in the realm of multi-view deep clustering, graph embedding methods have not received the same level of attention as in the single-view domain, leading to a lack of focus on local sample information and, consequently, hindering the enhancement of clustering performance. Therefore, this paper introduces a novel graph embedding-based deep multi-view clustering algorithm. This algorithm employs a co-training strategy to sustain the execution of the global clustering task and introduces an innovative graph embedding technique, providing additional prior information on the local similarity among sample features for the clustering task, thereby improving clustering outcomes. Extensive experiments on real-world datasets demonstrate that our algorithm outperforms current popular multi-view clustering algorithms.