Straightforward tensorized anchor graph learning for multi-view clustering

Abstract

Even while graph-based multi-view clustering methods are quite effective in capturing the connection between data and clustering structures, the majority of them still exhibit the following limitations: (1) some methods fail to account for higher-order correlations and spatial structures between multi-view data; (2) random sampling and k-Means lead to unstable selection of anchors; (3) post-processing steps in many studies contribute to suboptimal clustering performance. To solve these issues, we propose a straightforward tensorized anchor graph learning method (STAGL) for multi-view clustering, which integrates the low-rank tensor learning and clustering into a unified framework. Specifically, we first employ a variance-based decorrelation strategy to select anchor points and construct an anchor graph for every view. Based on this, STAGL explores the similarities and spatial structures of each view by minimizing the tensor-adaptive log-determinant regularization. Additionally, we directly employ the anchor graphs to obtain the final clustering assignments by computing the distances between samples. Meanwhile, an adaptive strategy is incorporated to account for the varying importance of different views in the clustering process. Finally, we employed an efficient algorithm to solve this model, and comprehensive experiments on six datasets demonstrate the superior clustering performance of the proposed method.