Anchor Graph Learning with Double Noise Removal for Multi-View Clustering.

Abstract

Existing anchor-based multi-view graph clustering methods have primarily focused on improving clustering performance through representative anchors selection. However, these methods fail to effectively eliminate noise across different views, resulting in inaccurate consensus representations that degrade clustering quality. To address this limitation, we propose a novel approach, termed Anchor Graph Learning with Double Noise Removal for Multi-view Clustering (AGLDR), which simultaneously learns a consistent anchor graph across views and removes view-specific noise from two distinct distributions. Specifically, we first introduce a low-rank constraint on the consistent anchor graph to capture global structural correlations in the data representation and enhance the model's robustness to noise. Subsequently, we minimize the F-norm and L_2,1 norm of two noise terms, which are designed to eliminate Gaussian-distributed and Laplacian-distributed noise, respectively. Our algorithm utilizes reasonable techniques to address an important gap in noise reduction in multi-view clustering. Extensive experiments demonstrate the superiority of AGLDR over state-of-the-art methods in terms of clustering accuracy and robustness. The code of this paper is available at https://github.com/chenzhe207/AGLDR.