Machine learning informed by micro- and mesoscopic statistical physics methods for community detection

Area of Science:

• Network Science
• Machine Learning
• Statistical Physics

Background:

• Community detection is vital for understanding complex network structures.
• Traditional methods often overlook fine-grained node similarities.
• Integrating micro-level similarities into mesoscopic structures remains a challenge.

Purpose of the Study:

• To propose a low-complexity framework integrating machine learning for enhanced community detection.
• To improve structural coherence and accuracy by embedding node-pair similarities.
• To outperform existing methods in identifying community structures.

Main Methods:

• Developed a framework embedding micro-level node-pair similarities into mesoscopic community structures.
• Utilized ensemble learning models to enhance detection.
• Evaluated performance on artificial and real-world networks.

Main Results:

• The proposed framework consistently outperforms conventional, embedding-based, and learning-based approaches.
• Achieved higher modularity, normalized mutual information, and adjusted rand index.
• Demonstrated significant accuracy improvements even without ground-truth information.

Conclusions:

• Machine learning enhances statistical physics methods for superior community detection.
• Node-pair similarity is critical for improving detection accuracy.
• The framework effectively uncovers complex structural patterns in networks.