An online Bayesian approach to change-point detection for categorical data

Change-point detection for categorical data has wide applications in many fields. Existing methods either are distribution-free, not utilizing categorical information sufficiently, or have limited performance when there exists “rare events” (events that occur with low frequency). In this paper, we propose a Bayesian change-point detection model for categorical data based on Dirichlet-multinomial mixtures. Because of the introduction of prior information, our method performs well for the existence of “rare events”. An online parameter estimation procedure and an online detection strategy are then designed to adapt to data streams. Monte Carlo simulations discuss the power of the proposed method and show advantages compared with existing algorithms. Applications in biomedical research, document analysis, health news case study and location monitoring indicate practical values of our method.