Reliable Multi-View Clustering with Graph Neural Network

Abstract

Deep multi-view clustering typically leverages deep neural networks to extract consistent information across multiple views. Graph neural networks gather significant attention due to its ability to model both topology and feature simultaneously, making them highly attractive for researchers working on multiview scenarios. However, due to the inherent heterogeneity of multi-view data, uncertainty may intrinsically reside within the data, and the message passing mechanism among samples can amplify such uncertainty embedded in the data. Existing multiview clustering methods based on graph neural networks fail to effectively address the above issues, leading to unreliable sample representation. To address the challenge above, we propose a model called Reliable Multi-View Clustering with Graph Neural Network (RM-GNN). RM-GNN first uses view-specific graph convolutional encoders to project features from different views into a unified semantic space. Then, we introduce an adaptive fusion method that effectively combines the representations obtained from each view.Due to the lack of supervisory information in deep multi-view clustering tasks, we propose a similarity-driven evidence learning module to effectively identify and distinguish unreliable samples in the embedding space. We model the similarity between the initial embedding and the consistent embedding by a Dirichlet distribution, enabling the quantification of subjective opinions regarding the reliability of embedding representations. This approach explicitly captures uncertainty within the embedding space. Based on both similarity and uncertainty, high-uncertainty samples are identified, and their influence on embedding reliability is mitigated. Additionally, we introduce a unified transformation matrix to enforce global topological consistency. Extensive experiments on eight benchmark datasets demonstrate that our algorithm consistently outperforms state-of-the-art methods.

Keywords: Deep Learning, Representation Learning, unsupervised learning, Multi-view Clustering, uncertainty estimation