GCN-based unsupervised community detection with refined structure centers and expanded pseudo-labeled set.

Area of Science:

• Graph theory
• Network analysis
• Machine learning

Background:

• Community detection is crucial for understanding graph-structured data.
• Graph convolutional networks (GCNs) offer unsupervised community detection but depend heavily on initial centers and have limited label propagation.
• Shallow GCNs struggle to disseminate limited label information across the entire graph due to localized filtering.

Purpose of the Study:

• To develop an improved GCN-based unsupervised community detection method (RE-GCN).
• To address the sensitivity of GCNs to initial structure centers.
• To enhance the label propagation capabilities of shallow GCNs for more effective community detection.

Main Methods:

• RE-GCN iteratively refines structure centers by alternating between GCN partitioning and updating centers based on subgraph importance.
• It expands the pseudo-labeled set by selecting nodes with affiliation strengths similar to their structure centers.
• The method integrates network topology and node attributes for community detection.

Main Results:

• The refinement process generates more representative structure centers.
• Expanding the pseudo-labeled set significantly improves GCN performance in community detection.
• RE-GCN demonstrates effectiveness on both attributed and non-attributed networks.

Conclusions:

• RE-GCN offers a robust approach to unsupervised community detection by refining structure centers and expanding pseudo-labeled data.
• The method overcomes limitations of traditional GCNs in handling initial center sensitivity and label propagation.
• The proposed technique enhances the accuracy and reliability of community detection in complex networks.