High-order correlation and consistency-aware multi-view clustering via anchor graph learning.

Multi-view clustering has become an effective tool for integrating complementary information from multiple data sources. However, traditional clustering methods often struggle with computational efficiency, as well as capturing high-order correlations across views and maintaining cross-view consistency, especially when dealing with large-scale and heterogeneous data. To handle these issues, we propose a novel framework, High-Order Correlation and Consistency-Aware Multi-View Clustering via Anchor Graph Learning (HCAGL). HCAGL leverages anchor graph learning using a compact set of anchor points, which effectively reduce dimensionality and substantially enhances computational efficiency. To capture high-order correlation among views, we introduce a tensor Schatten p-norm on the low-dimensional anchor embeddings, facilitating the propagation of global consistency across views. Furthermore, we incorporate an adaptive neighborhood graph learning strategy to construct a consensus graph, dynamically adjusting view-specific weights based on their relative importance, thereby enhancing cross-view consistency. Extensive experiments on six benchmark datasets demonstrate the superior performance of HCAGL in capturing cross-view consistency and high-order correlations, achieving higher accuracy and better clustering quality than existing multi-view methods. Analyses of model components and parameter sensitivity show that each design choice contributes positively to performance, ensuring stable and reliable results across diverse datasets. These findings indicate that HCAGL is an effective and computationally efficient solution for complex multi-view clustering problems.