Abstract:

Geospatial data mining aims to deeply reveal the complex distribution rules and spatio-temporal evolution trends of multiple geographic elements. Current geospatial data mining studies were mostly based on the assumption of spatial correlation dependence, which lacked the analysis of underlying spatial causalities, so the mixed pseudo-correlations would lead to biased or even erroneous mining results. In this case, based on the causal inference theory, this study proposed a multivariate geographic element causal pattern mining method by considering the influence of spatial neighborhood effects on causal relationships. First, the transaction set adapted to the distribution density of geographic elements is automatically created using the spatial clustering algorithm. Then, the spatial causal directed graph structure of multiple geographic elements is constructed by integrating the spatial neighborhood effects and the Bayesian network modeling idea. Finally, the backdoor criterion is used to implement the intervention operation to realize the quantitative calculation of causal effects among multiple geographic elements. In the experiments, the spatial distribution data of urban facilities in Shenzhen and Shanghai are utilized for case studies. The comparison results with the spatial correlation pattern mining method show that the proposed method eliminates the spurious spatial correlations caused by confounding variables, and can effectively obtain the directed causal relationships and causal strength between different types of urban functional facilities, and reveal the local aggregation effects of urban functional facilities more accurately. The mined causal patterns have the potential of providing more credible decision supports for the optimization of urban space layout.

Key words: causal patterns, multiple geographic elements, spatial neighborhoods, causal effects