Monitoring binary networks for anomalous communication patterns based on the structural statistics

Network monitoring has become increasingly popular in the area of statistical process control due to its wide applications in fraud detection, corporate management and political behavioral analysis. This paper focuses on the cases where the communication pattern is significant while changes in specific nodes are negligible. The important features including the density, reciprocity, degree variability, and transitivity are considered to reflect the commonly-encountered communication patterns in social networks. The structural statistics are provided for characterizing the main features. A multivariate control chart is adopted to monitor the structural statistics simultaneously so as to account for their correlations and to decrease the overall false alarm rate. A performance evaluation framework is proposed based on the Exponential Random Graph Models (ERGMs) in order to simulate the shifts of communication patterns. The results of the numerical experiments show that the Hotelling T² control chart for the structural statistics outperforms several benchmark methods especially in detecting the large shifts of reciprocity and transitivity. The effectiveness of the proposed method is validated through the analysis of the Enron email communication networks.