Detecting Multiple Changepoints by Exploiting Their Spatiotemporal Correlations: A Bayesian Hierarchical Approach

Capturing the nonstationarity of spatiotemporal data over time via changepoints has received increasing attention in various research fields. Although extensive studies have been conducted to investigate changepoint detection with spatiotemporal data, research on detecting multiple clusters of spatiotemporally correlated changepoints has remained unexplored. In this paper, we propose a multilayer Bayesian hierarchical model: The first layer uncovers the spatiotemporal correlations of changepoints based on multiple propagation binary variables, which describe the occurrences of change propagations. The second and third layers compose nonhomogeneous hidden Markov models to capture time series data and their state sequences, in which changes of states signify changepoints. We perform Bayesian inference for changepoints and change propagations via a forward-backward algorithm that combines recursion and Gibbs sampling. Based on the experiments with simulated data, we show that our method significantly improves the detection accuracy toward spatiotemporally correlated changepoints. A real-world application to bike-sharing data also demonstrates the effectiveness of our method. This research has significant relevance to companies operating systems across geographical regions, as it enables a more robust understanding of emerging trends and shifts in spatiotemporal data.