Auto-weighted multi-dimensional feature fusion for incomplete multi-view clustering

Multi-view subspace clustering is an effective method for clustering high-dimensional data but faces several limitations: (1) It often clusters high-dimensional data directly, overlooking the redundancy of original features and the relevance of features across different dimensions. (2) Higher-order correlations and differential structures between views are frequently ignored, leading to suboptimal performance of the fused subspace representation matrix. To address these issues, we propose an auto-weighted multi-dimensional feature fusion incomplete multi-view clustering method (AWMDF${}^2$). AWMDF${}^2$ enhances data representation by decomposing the completed data kernel matrix into feature matrices of various dimensions, which are then automatically weighted according to their contribution. These weighted matrices are fused into a consensus feature matrix, which replaces the original high-dimensional data for subspace learning. Additionally, we develop a multi-view subspace fusion method based on the weighted tensor Schatten-p norm, which captures higher-order relationships between views and assigns appropriate weights to each view. AWMDF${}^2$ integrates multi-dimensional feature fusion, subspace learning, and higher-order relational learning into a unified optimization framework. Extensive experiments on six public datasets demonstrate that AWMDF${}^2$ outperforms ten existing advanced baseline methods.