Cluster-graph convolution networks for robust multi-view clustering

Abstract

Existing deep contrastive representation learning methods for unlabeled multi-view data have shown impressive performance by shrinking the cross-view discrepancy. However, most of these methods primarily focus on the procedure of common semantics extraction from multiple views, which is just one of the factors affecting the performance of unsupervised multi-view representation learning. Two additional factors are often overlooked: i) how to improve the discriminative ability of final representations. Existing unsupervised-based approaches normally perform worse on clustering as the number of categories increases. ii) how to balance the contribution of multiple views (specifically in data with more than two views). We observe that the quality of the learned representation is also influenced by certain views, i.e., the model precision may be decreased when some views are involved in the training. To address these factors, we propose a novel contrastive learning-based method, called Cluster-Graph Convolution networks for Robust Multi-view Clustering (CGC-RMC), for unlabeled multi-view data. Specifically, we design a specialized spatial-based cluster-graph convolution and a new adaptive sample-weighted strategy in a contrastive-based basic framework for the above two factors. Additionally, the proposed method adopts a communication fusion module to relieve the influence of view-private information in final view representations. Extensive experiments demonstrate that the proposed method outperforms eleven competitive unsupervised representation learning methods on six multi-view datasets based on the performance of the learned representation on the clustering task.