Redundancy-Free Self-Supervised Relational Learning for Graph Clustering.

Area of Science:

• Data Science
• Machine Learning
• Network Analysis

Background:

• Graph clustering is crucial for data analysis, with Graph Neural Networks (GNNs) gaining traction.
• Existing methods often neglect relational information between non-independent, non-identically distributed nodes.
• This oversight limits the exploitation of semantic information, leading to suboptimal clustering.

Purpose of the Study:

• To propose a novel self-supervised deep graph clustering method, Relational Redundancy-Free Graph Clustering (R2FGC).
• To address the limitations of existing methods by effectively utilizing attribute- and structure-level relational information.
• To improve the performance of graph clustering by learning discriminative node embeddings.

Main Methods:

• R2FGC employs an autoencoder (AE) and a graph AE (GAE) to extract relational information from global and local views.
• It preserves consistent relationships among augmented nodes to capture semantic information.
• Redundant relationships are reduced, and a strategy is implemented to mitigate the oversmoothing issue.

Main Results:

• R2FGC demonstrates superior performance compared to state-of-the-art baselines on benchmark datasets.
• The method effectively extracts attribute- and structure-level relational information.
• Learned embeddings are more discriminative, leading to enhanced cluster assignments.

Conclusions:

• R2FGC offers a significant advancement in self-supervised deep graph clustering.
• The approach effectively leverages relational information for improved clustering accuracy.
• The method provides a valuable tool for analyzing complex graph-structured data.