Bag of DAGs: Flexible & Scalable Modeling of Spatiotemporal Dependence

We propose a computationally efficient approach to construct a class of nonstationary spatiotemporal processes in point-referenced geostatistical models. Current methods that impose nonstationarity directly on covariance functions of Gaussian processes (GPs) often suffer from computational bottlenecks, causing researchers to choose less appropriate alternatives in many applications. A main contribution of this paper is the development of a well-defined nonstationary process using multiple yet simple directed acyclic graphs (DAGs), which leads to computational efficiency, flexibility, and interpretability. Rather than acting on the covariance functions, we induce nonstationarity via sparse DAGs across domain partitions, whose edges are interpreted as directional correlation patterns in space and time. We account for uncertainty about these patterns by considering local mixtures of DAGs, leading to a ``bag of DAGs'' approach. We are motivated by spatiotemporal modeling of air pollutants in which a directed edge in DAGs represents a prevailing wind direction causing some associated covariance in the pollutants; for example, an edge for northwest to southeast winds. We establish Bayesian hierarchical models embedding the resulting nonstationary process from the bag of DAGs approach and illustrate inferential and performance gains of the methods compared to existing alternatives. We consider a novel application focusing on the analysis of fine particulate matter (PM2.5) in South Korea and the United States. The code for all analyses is publicly available on Github.