Robust Multi-View Clustering via Graph-Oriented High-Order Correlations Learning

Abstract

Multi-view clustering aims to partition data into corresponding clusters by leveraging features from various views to reveal the underlying structure of the data fully. However, existing multi-view clustering methods, particularly graph-based techniques, face two main issues: 1) They often construct similarity matrices directly from low-quality and inflexible graphs, resulting in inadequate fusion of multi-view information and impacting clustering performance; 2) Most methods focus only on consensus or pairwise associations between views, neglecting more complex higher-order correlations among multiple views, which limits improvements in clustering performance. To address these issues, we propose a novel multi-view clustering method called Robust Multi-View Clustering via Graph-Oriented High-Order Correlations Learning (GHCL). GHCL first learns latent embeddings for each view and stacks these embeddings into a third-order tensor. Then, Tucker decomposition and regularization constraints are applied to optimize the tensor and error terms, producing high-quality denoised graphs. Additionally, GHCL introduces an adaptive confidence mechanism that integrates the learned similarity matrix and consensus representation into a unified step, enhancing multi-view information fusion and clustering effectiveness. Extensive experiments demonstrate that GHCL significantly outperforms current state-of-the-art techniques on multiple datasets. It effectively integrates multi-view information and captures higher-order correlations between views, improving clustering accuracy and robustness in handling complex data, thereby showcasing its practical value in multi-view data analysis.