ComClus: A Self-Grouping Framework for Multi-Network Clustering.

Area of Science:

• Network analysis
• Machine learning
• Data mining

Background:

• Multi-network clustering assumes a shared structure, which is often too strict for diverse real-world data.
• Existing methods struggle when networks belong to different underlying groups with distinct clustering patterns.

Purpose of the Study:

• To develop a method that simultaneously groups and clusters multiple networks.
• To address the limitation of diverse data distributions across networks.
• To improve clustering accuracy by accounting for underlying network groups.

Main Methods:

• Proposes ComClus, a novel method combining non-negative matrix factorization (NMF) for clustering and metric learning for feature subspace learning.
• Treats node clusters as network features to learn subspaces differentiating network groups.
• Couples network grouping and clustering for mutual enhancement.
• Incorporates semi-supervised learning to leverage prior knowledge for network grouping.

Main Results:

• ComClus effectively detects network groups and clusters networks within these groups.
• The coupled approach of grouping and clustering enhances overall accuracy.
• Semi-supervised grouping further boosts clustering performance.
• Demonstrates effectiveness and scalability on synthetic and real-world datasets.

Conclusions:

• ComClus offers an effective solution for joint clustering of multiple networks with diverse structures.
• The method's ability to group networks and leverage prior knowledge improves clustering accuracy.
• ComClus is a scalable and robust approach for complex network analysis.