Tensor-Based Adaptive Consensus Graph Learning for Multi-View Clustering

Multi-view clustering has garnered considerable attention in recent years owing to its impressive performance in processing high-dimensional data. Most multi-view clustering models still encounter the following limitations. They emphasize common representations or pairwise correlations between multiple views, while neglecting high-order correlations. The weights of multiple views or prior information of singular values are ignored in the clustering process. Therefore, a Tensor-based Adaptive Consensus Graph Learning (TACGL) model is proposed for addressing above problems. Specifically, all representation matrices of multiple views are stacked into a representation tensor to reveal high-order connections among multiple views. A weighted tensor nuclear norm is imposed on representation tensor to maintain property of low-rank and discovers the prior information of singular values. The weights of graph learning can be automatically assigned to each similarity graph via consensus graph learning, resulting in a unified graph matrix. Laplacian rank constraint is imposed on the unified matrix to help partition the samples into the desired number of clusters. An algorithm based on Alternating Direction Method of Multipliers (ADMM) is designed for solving TACGL. Based on comprehensive experiments conducted on ten datasets, it is clear that the proposed model showcases substantial advantages over fourteen state-of-the-art models.