Modularity-aware graph autoencoders for joint community detection and link prediction.

Area of Science:

• Graph Machine Learning
• Network Science
• Data Mining

Background:

• Graph Autoencoders (GAE) and Variational Graph Autoencoders (VGAE) excel at link prediction but underperform in community detection compared to simpler methods like Louvain.
• The potential for GAE/VGAE to enhance community detection, particularly without node features, and maintain link prediction accuracy remains uncertain.

Purpose of the Study:

• To investigate the joint optimization of link prediction and community detection using GAE and VGAE.
• To develop an improved GAE/VGAE framework that enhances community detection accuracy without sacrificing link prediction performance.

Main Methods:

• Introduced a community-preserving message passing scheme for GAE and VGAE encoders.
• Incorporated graph structure and modularity-based prior communities into embedding space computation.
• Proposed novel training strategies, including a modularity-inspired regularizer, complementing reconstruction losses.

Main Results:

• Demonstrated the possibility of jointly achieving high accuracy in both link prediction and community detection.
• Empirically validated the effectiveness of the proposed Modularity-Aware GAE and VGAE approach.
• Showcased improved performance on various real-world graph datasets.

Conclusions:

• Modularity-Aware GAE and VGAE effectively address the dual tasks of link prediction and community detection.
• The proposed methods offer a significant advancement over existing GAE/VGAE techniques for community detection.
• This work provides a robust framework for analyzing graph data with improved accuracy in both link prediction and community detection.