Detecting spatial flow outliers in the presence of spatial autocorrelation

Spatial flow outlier (SFO) detection aims to discover spatial flows whose non-spatial attribute values are significantly different from their neighborhoods. Different from spatial flow clusters, which are the main concern in the current literature, SFOs represent unusual local instabilities and are valuable for revealing anomalous spatial interactions between regions. Detecting SFOs is challenging because the underlying distribution of the flow data is unknown a priori, and inappropriate distribution assumptions may lead to misleading decisions on SFOs. Surprisingly, spatial autocorrelation, which is a common property of geographic data, has not been considered in the null hypothesis for testing spatial outliers. To solve this significant methodological issue, we propose a spatial-autocorrelation-aware detection method. This method detects SFOs by testing the local difference of attribute values in flow neighborhoods against the null hypothesis that neighboring flows are similar. To construct this null hypothesis, we develop a distribution-free model by reconstructing the observed spatial autocorrelation. Synthetic experiments and a case study using the journey-to-work flow data in Chicago demonstrate that the choice and modeling of the null hypothesis has a significant influence on the statistical inference of SFOs. By taking the inherent spatial autocorrelation into account, our method can more objectively assess the significance of SFOs than two baseline methods based on the normality and randomization hypotheses.