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

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph convolutional networks (GCNs) show promise in graph clustering but struggle with multiview data.
• Multiview representation learning offers rich information but presents challenges for unified GCN processing.

Purpose of the Study:

• To develop a novel multiview graph clustering strategy integrating contrastive learning (CL) and reinforcement learning (RL).
• To enhance the modeling and processing of complex, diverse features in multiview graph data.

Main Methods:

• Constructed local and global information views, capturing features via an adaptive neighborhood-aware algorithm.
• Employed a channel attention mechanism for weighted aggregation of neighborhood information.
• Utilized CL for feature differentiation (enhancement, masking) and integrated semantic/structural node information.
• Designed a reinforcement learning (RL) based rewriter module using proximal policy optimization (PPO) for fine-grained cluster boundary adjustments.

Main Results:

• The proposed method demonstrated superior clustering accuracy compared to ten state-of-the-art algorithms.
• Achieved enhanced model robustness and reliability in graph clustering tasks.
• Effectively refined graph partitioning through RL-guided fine-grained adjustments.

Conclusions:

• The proposed multiview CL and RL strategy offers a significant advancement in graph clustering.
• The coarse-to-fine fusion approach effectively handles complex multiview graph data.
• The method shows strong potential for real-world graph clustering applications.