A Deviation-Based Framework for Unified Community and Anomaly Detection in Social Networks

Community detection and anomaly detection are fundamental tasks in social network analysis. While communities represent cohesive groups of users with similar interaction patterns, anomalies are nodes whose behavior deviates from established norms. Traditional methods often address these tasks separately or rely on similarity- and embedding-based measures, which can limit both interpretability and scalability. This work proposes a Deviation-Based model that unifies community and anomaly detection within a single framework. Communities are initialized from high-degree seed nodes and expanded iteratively, whereas anomalies naturally emerge as statistical outliers that deviate from structural and attribute distributions. Experiments on benchmark citation networks, real-world social platforms, and synthetic graphs demonstrate that the proposed model outperforms both classical structural and deep learning-based baselines in terms of community detection (Normalized Mutual Information (NMI), Adjusted Rand Index (ARI), modularity) and anomaly detection (precision, recall, F1-score, Area Under the Curve (AUC)). The model is lightweight, scalable, and interpretable, offering a practical solution for large-scale social networks and applications in security, fraud detection, and information management.