Anomaly Detection using Adaptive Fusion of Graph Features on a Time Series of Graphs

Abstract It is known that fusion of information from graph features, compared to individual features, can provide superior inference for anomaly detection [PPM10]. However, selection of a fusion technique other than a naive equal weighting is not trivial. We present a multivariate methodology for fusion of features derived from time series of graphs, and investigate its inferential efficacy. The results demonstrate that our methodology has higher detection, estimation, and localization power than the standard linear weighting fusion techniques for a certain class of anomaly detection problems. Simulation results using a time series of random graphs, as well as illustrative results from a time series of Enron email data, are presented.