Change-Point Detection in Time Series of Weighted Stochastic Block Model Graphs

• Activity Number: 440 - SLDS CSpeed 8
• Type: Contributed
• Date/Time: Thursday, August 12, 2021 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Carolyn Liu* and Heng Wang and Youngser Park and Carey E Priebe
• Companies: Ward Melville High School and AppLovin and Johns Hopkins University and Johns Hopkins University
• Keywords: Anomaly Detection; Graph Mining; Temporal Graphs; Scan Statistics

Abstract

Dynamic network data can be readily observed as a time series of graphs. Entities are represented by vertices and time-evolving edges are denoted by the relationships between these entities, which can be a natural framework for inquiry. Nevertheless, within the emerging study of dynamic graph mining, the change-point detection problem is one that has become progressively prominent. Using the hypothesis-testing method in weighted stochastic block model time series, we study the power characteristics of two competing scan statistics built on distinct underlying locality statistics. Moreover, we examine the application of the two test statistics on real world temporal graph data. Furthermore, we compare these two scan statistics directed under a weighting schema in a stochastic block model time series in terms of statistical power dominance and robustness. These results allow for the investigation of more general stochastic block model time series graphs with weighting assumptions.