Bayesian Network Scan Statistics for Multivariate Pattern Detection

Abstract

We review three recently proposed scan statistic methods for multivariate pattern detection. Each method models the relationship between multiple observed and hidden variables using a Bayesian network structure, drawing inferences about the underlying pattern type and the affected subset of the data. We first discuss the multivariate Bayesian scan statistic (MBSS) proposed by Neill and Cooper (2008). MBSS is a stream-based event surveillance framework that detects and characterizes events given the aggregate counts for multiple data streams. Next, we describe the agent-based Bayesian scan statistic (ABSS) proposed by Jiang et al. (2008). ABSS performs event detection and characterization given individual-level data for each agent in a population. Finally, we review the anomalous group detection (AGD) method proposed by Das, Schneider, and Neill (2008). AGD is a general pattern detection approach which learns a Bayesian network structure from data and detects anomalous groups of records.