Collaborative Decision-Reinforced Self-Supervision for Attributed Graph Clustering

Area of Science:

• Graph theory
• Machine learning
• Data mining

Background:

• Attributed graph clustering partitions graph nodes into groups.
• Variational graph autoencoders (VGAE) are commonly used for node representation learning.
• Integrating supervised information to enhance graph node representation learning for clustering remains a challenge.

Purpose of the Study:

• To propose a novel method, Collaborative Decision-Reinforced Self-Supervision (CDRS), for attributed graph clustering.
• To enhance node representation learning by integrating supervised information into existing VGAE-based methods.
• To improve graph clustering performance through a collaborative, self-supervised approach.

Main Methods:

• Introduced a transformation module for end-to-end training of VGAE-based methods.
• Implemented a pseudo node classification task via multitask learning.
• Developed a self-supervision strategy using a dynamically augmented pseudo-label set based on consistent clustering and classification decisions.
• Investigated sorting strategies to optimize the pseudo-label set quality.

Main Results:

• The proposed CDRS method demonstrated superior performance compared to state-of-the-art methods across multiple datasets.
• The collaborative decision-making process effectively enhanced node representation learning.
• The self-supervision mechanism progressively improved the network's generalization capability.

Conclusions:

• CDRS offers a robust framework for attributed graph clustering by effectively leveraging self-supervision.
• The integration of pseudo node classification significantly boosts clustering performance.
• The method provides a promising direction for incorporating supervised information in graph representation learning.