Anomaly Detection in Time-Evolving Networks Using Tensor Spectrum

• Activity Number: 498
• Type: Contributed
• Date/Time: Wednesday, August 3, 2016 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Ruikai Cao* and Yulia R. Gel
• Companies: The University of Texas at Dallas and The University of Texas at Dallas
• Keywords: dynamic networks ; anomaly detection ; tensor spectrum ; higer order motifs ; temporal dependency ; Enron emails

Abstract

In analysis of dynamic networks, one of the key tasks are anomaly detection. Its applications range from new gang formation to brain damages to money laundering. Most of the currently available methods for anomaly detection have two disadvantages: either they focus only on two-dimensional structures, that is, edges connecting pairs of nodes; or they neglect the important temporal dependence structure of change point statistics in networks, which in turn leads to distorted false positive and false negative rates. In this talk we circumvent these problems by introducing a new anomaly detection method based on tensor spectral characteristics. The new data-driven approach is distribution-free and allows to detect change points in higher-order network motifs. We evaluate our new anomaly detection procedure on synthetic networks and the Enron communication benchmark case study.