Anchor-based scalable multi-view subspace clustering

Abstract

Multi-view clustering is a major topic in pattern recognition and machine learning. Common multi-view clustering algorithms construct similarity graphs from original samples and use them to perform spectral clustering. The time complexity of the singular value decomposition process in graph construction and spectral clustering is high, leading to high computational and memory costs. In addition, subsequent K-means clustering is sensitive to the initial points, yielding unstable clustering results. To address these issues, this study proposes a novel approach to reduce time overhead and memory space from two perspectives. First, a new anchor selection method is proposed to reduce the dimension of the original data by lowering the cost. Second, the self-representation matrix of multi-views is fused into a consistent graph matrix using the post-fusion technique, and the fused graph is directly processed in postprocessing. Furthermore, the proposed method directly obtains clustering results based on the connectivity of the fusion graph, eliminating the need for K-means postprocessing, which avoids the issue of unstable clustering results. Experimental results on artificial and real multi-view datasets indicate that the proposed algorithm is superior to existing algorithms.