Co-Membership-based Generic Anomalous Communities Detection

Area of Science:

• Network Science
• Data Mining
• Machine Learning

Background:

• Detecting anomalous communities in networks is vital for uncovering network insights.
• Existing methods often rely on vertex attributes or community topology.
• A gap exists in methods utilizing vertex co-membership across multiple communities.

Purpose of the Study:

• Introduce a novel, generic algorithm for anomalous community detection.
• Develop a method robust to community size, density, and domain.
• Provide a tool for generating labeled anomaly-infused networks for research.

Main Methods:

• Propose the Co-Membership-based Generic Anomalous Communities Detection Algorithm (CMMAC).
• Utilize vertex co-membership information across multiple communities.
• Train a classifier to predict vertex community membership probability and rank communities.

Main Results:

• CMMAC demonstrates superior performance over existing methods on simulated and real-world networks.
• The algorithm is domain-free and resilient to variations in community size and density.
• Successfully identified abnormal communities in unlabeled real-world networks (e.g., Reddit, Wikipedia).

Conclusions:

• CMMAC offers a powerful and versatile approach for anomalous community detection.
• The developed network generation algorithm facilitates reproducible research in this field.
• CMMAC shows practical applicability across diverse network domains.