Coarse-to-Fine Fusion: Customized Multiview Contrast Reinforcement Learning for Graph Clustering

The application of graph convolutional networks (GCNs) in the field of graph clustering has achieved remarkable results. However, there still exist certain limitations. Multiview representation learning is suitable for graph clustering tasks because it contains rich information. Nevertheless, when dealing with multiview data, GCNs usually have difficulty effectively modeling and processing its complex and diverse features in a unified manner. To address the above challenges, a multiview contrastive learning (CL) and reinforcement learning (RL) strategy for graph clustering with coarse-to-fine fusion is proposed in this article. First, a view centered on local information and a view centered on global information are constructed, respectively. The feature information of the two views is captured by an adaptive neighborhood-aware algorithm, and a graph representation containing information from different views is learned. The channel attention mechanism is introduced to select key neighborhood information from different feature channels of each view and perform weighted aggregation based on the learned weights. Then, a CL strategy is used to differentiate the features of different views, including feature enhancement and feature masking, in order to enhance the model's ability to recognize complex graph structures. Simultaneously integrating the dual information of semantics and structure from nodes across different views to enhance model robustness and the reliability of graph clustering. Finally, a rewriter module based on improved proximal policy optimization (PPO) is designed. Leveraging the intelligent decision-making mechanism of RL, this module introduces a strategy constraint term tailored for clustering tasks into the standard PPO framework. Guided by this constraint, the agent performs fine-grained adjustments and summarization of nodes at cluster boundaries, thereby further refining graph partitioning results and enhancing the accuracy of graph clustering. Comparative experimental analysis with ten state-of-the-art graph clustering algorithms on six real-world datasets shows that the clustering accuracy of this algorithm is better.