A Nonparametric Test-Based Approach for Mining Spatio-Temporal Co-Occurrence Patterns of Urban Crimes

Scientific suggestions for crime prevention and control can be provided by analyzing the association relationship among multi-types of crimes based on spatio-temporal co-occurrence pattern discovery method. User-specified thresholds of prevalence measures are usually required by existing methods to filter mining results. Wrong decisions may be made by application departments without enough prior knowledge. Thus, a significance test method is proposed for mining spatio-temporal co-occurrence patterns among urban crimes. Firstly, a spatio-temporal pattern reconstruction method is developed to construct the null model of independence by fitting the observed distribution characteristics of each feature. Then, the significance of candidate spatio-temporal co-occurrence patterns are tested based on the empirical distributions of co-occurrence prevalence of candidate patterns under the null model. Simulated datasets with predefined patterns are further used to verify the effectiveness of this method. In addition, the spatio-temporal cooccurrence patterns among 13 types of crimes of the city S in 2016 are identified at multiple analysis scales (i.e. spatio-temporal radius). Taking the pattern disorderly conduct, motor vehicle theft, pickpocketing as an example, the formation mechanisms of that pattern are deeply analyzed by combining with the spatial distributions of communal facilities. The result shows that:(1) statistically significant spatio-temporal cooccurrence patterns can be effectively detected by fully considering the effect of autocorrelation of each type of crime; (2) spatio-temporal co-occurrence patterns among crimes vary with the scales of analysis; and (3) spatio-temporal co-occurrence patterns usually happen among different crimes with similar artificial and social environment.