A Bayesian Markov Switching Model for Change Point Detection in Sparse Causal Graphical Learning

• Activity Number: 247
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 2:00 PM to 3:50 PM
• Sponsor: IMS
• Author(s): Huijing Jiang*+ and Fei Liu and Aurelie Lozano
• Companies: IBM Thomas J. Watson Research Center and IBM Thomas J. Watson Research Center and IBM Thomas J. Watson Research Center
• Keywords: Causal inference ; Sparse graphical learning ; Multivariate time series ; Markov switching model ; Group variable selection ; Change point detection

Abstract

Causal inference is an important topic in statistics and machine learning and has wide applicability ranging from biology to social sciences. Learning temporal graphical structures from multivariate time series data reveals important dependency relationship between current and past observations and is thus a key research focus for causal discovery. Most of the traditional methods assume a "static" temporal graph. Yet in many relevant applications, the underlying dependency structures may vary over time. In addition, with particular focus on the sparsity of the resulting causal graphical models, the lagged variables belonging to the same time series shall be included or excluded simultaneously. In this paper, we introduce a Markov switching vector autoregressive model to detect the structural changes of the causal relationship in multivariate time-series data. Our approach allows for such structural changes by a set of latent state variables modeled by a Markov process. At each state, we further impose the sparse structure of the causal graphical models through the hierarchical Bayesian group Lasso method. We demonstrate the value of our approach on simulated and real-world datasets