|
Abstract:
|
Dynamic network models can be used to represent the constantly changing nature of the interactions and relationships between objects or people. This paper focuses on network anomaly detection, the process of discovering unusual times or structures, in large dynamic networks. We propose a two-stage approach that can quickly detect anomalous vertices. In the first stage, a discrepancy score is developed that measures the unusualness of recent edges. The second stage applies multivariate change point detection methods on the discrepancy measures to identify the edges and vertices that have experienced a change. We validate our method using simulated data and real publicly available data with known change points of different types.
|
