Beyond Stationarity in Time Series: Discovering Causal Structures and Latent Regimes via Markov Blankets

To address the limitations of traditional time series causal discovery methods, which assume a single, consistent causal structure, researchers have proposed a new algorithm called RCBNB-MB. This algorithm identifies a subset of time points containing stable causal structures (i.e., potential causal regimes) and uses an iterative strategy to segment the time series, discovering causal graphs independently within each segment. Unlike directly using parent nodes, RCBNB-MB employs Markov blankets for inference, aiming to enhance robustness against causal discovery errors and retain predictive information. The authors provide theoretical guarantees regarding regime transitions and causal graphs under reasonable assumptions. The algorithm has been widely validated on simulated datasets with known true values and real-time IT monitoring data where regime transitions are critical to consider, with results showing that it systematically outperforms baseline methods in accurately detecting regime changes and their associated causal graphs.