Causal Structure Discovery for Spatio-temporal Data

Abstract

Numerous causal structure discovery methods have been proposed recently but none of them has taken possible time-varying structure into consideration. In this paper, we introduce a notion of causal time-varying dynamic Bayesian network (CTV-DBN) and define a causal boundary to govern cross time information sharing. Although spatio-temporal data have been investigated by multiple disciplines; by reducing structure discovery into a set of optimization problems, CTV-DBN is a scalable solution targeting large datasets. CTV-DBN is constructed using asymmetric kernels to address sample scarcity and to adhere to causal principles; while maintaining good variance and bias trade-off. We explore trajectory data collected from mobile devices which are known to exhibit heterogeneous patterns, data sparseness and distribution skewness. Contrary to a naïve method to divide space by grids, we capture the moving objects’ view of space by using density-based clustering to overcome the problems. In our experiments, CTV-DBN is used to reveal the evolution of time-varying region macro structure in a ring road system based on trajectories, and to obtain a local time-varying road junction dependency structure based on static traffic flow sensor data.