Extracting Causal Rules from Spatio-Temporal Data

This paper is concerned with the problem of detecting causality in spatiotemporal data. In contrast to most previous work on causality, we adopt a logical rather than a probabilistic approach. By defining the logical form of the desired causal rules, the algorithm developed in this paper searches for instances of rules of that form that explain as fully as possible the observations found in a data set. Experiments with synthetic data, where the underlying causal rules are known, show that in many cases the algorithm is able to retrieve close approximations to the rules that generated the data. However, experiments with real data concerning the movement of fish in a large Australian river system reveal significant practical limitations, primarily as a consequence of the coarse granularity of such movement data. In response, instead of focusing on strict causation (where an environmental event initiates a movement event), further experiments focused on perpetuation (where environmental conditions are the drivers of ongoing processes of movement). After retasking to search for a different logical form of rules compatible with perpetuation, our algorithm was able to identify perpetuation rules that explain a significant proportion of the fish movements. For example, approximately one fifth of the detected long-range movements of fish over a period of six years were accounted for by 26 rules taking account of variations in water-level alone.