Graph Joint Representation Clustering via Penalized Graph Contrastive Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Graph clustering is crucial for analyzing complex network data.
• Graph contrastive learning (GCL) is a dominant paradigm but suffers from false negative samples, hindering performance.
• False negative samples distort learned representations and limit clustering accuracy.

Purpose of the Study:

• To propose a graph clustering method that mitigates the impact of false negative samples in GCL.
• To improve the quality of learned representations by maintaining mutual information with input data.
• To enhance overall graph clustering performance.

Main Methods:

• Proposed maintaining mutual information (MI) between representations and inputs to reduce semantic loss from false negatives.
• Developed a GCL method penalized by reconstruction error, approximating MI maximization.
• Designed a specialized reconstruction decoder and error term to boost clustering.
• Incorporated a pseudo-label-guided strategy to further refine the GCL process.

Main Results:

• Experimental validation confirmed the effectiveness of maintaining MI.
• The proposed GCL method with reconstruction error penalty demonstrated improved clustering performance.
• The pseudo-label-guided strategy further alleviated issues caused by false negative samples.
• The new method outperformed state-of-the-art graph clustering algorithms.

Conclusions:

• The proposed graph clustering method effectively addresses the challenge of false negative samples in GCL.
• Integrating reconstruction error and pseudo-label guidance offers a promising direction for advanced graph clustering.
• The method shows significant potential for real-world applications requiring accurate graph analysis.