Multi-view dual anchor graph fuzzy clustering

Abstract

In recent years, with the rapid development of multi-view learning, how to effectively integrate information from different views for clustering analysis has become an important research topic in both academia and industry,driving the emergence of numerous efficient methods. However, current methods still face three major challenges.First, suboptimal anchor graphs often result from the inherent uncertainty and low discriminability of real-world data.Second, prevalent approaches primarily focus on common information between views, overlooking valuable view-specific information. Third, effectively leveraging the learned anchor graph to improve clustering remains under-explored.To overcome these challenges, this paper proposes a novel dual anchor graph fuzzy clustering framework. To address the first two challenges, we design a matrix factorization-based dual anchor graph learning framework. This framework extracts discriminative hidden representations from each view and subsequently derives both common and specific anchor graphs. For the third challenge, we develop an anchor graph fuzzy clustering method with a cooperative learning mechanism. This method constructs a dual anchor graph-driven fuzzy membership structure preservation mechanism to enhance clustering quality. Additionally, we introduce negative Shannon entropy to achieve adaptive view weighting. Extensive experiments on multiple benchmark datasets validate the effectiveness of the proposed DAG＿FC method. The results show that DAG＿FC outperforms competing methods on most metrics and datasets,achieving NMI improvements of approximately 30% and 20% over comparative methods on the Yale dataset. Moreover, the experiments also confirm that anchor graph-based clustering methods generally perform better than traditional subspace-based clustering methods. By incorporating hidden representation extraction techniques and designing specialized clustering algorithms, this paper further enhances the clustering performance of the proposed method.