Table 1.
Record fields, open police data [42].
Crime types are listed in S2 Appendix.
Fig 1.
Histogram of distances from anonymised to true crime locations (Cardiff).
Fig 2.
Hourly reported violence count by day of week (Cardiff).
Fig 3.
Subset of point of interest hierarchy.
Fig 4.
City centres used for evaluation.
Blue squares indicate 3 km × 3 evaluation areas. © OpenStreetMap [44] contributors.
Table 2.
Cities in the study and their populations, 2011 UK Census.
Note that population counts refer to entire city, Nviolence and NPOI are the violence and POI counts in the central 9 km2 of these cities.
Fig 5.
Illustration of regression from POI densities to violent crime density.
Fig 6.
Clustered correlation coefficients between subset of pairs of POIs (Cardiff).
Fig 7.
Model performance and variable count by LASSOα value, all cities.
Fig 8.
Model performance by kernel type, brackets show mean and standard deviation in performance.
Fig 9.
The effect of varying the bandwidth in POI KDE, by city.
Fig 10.
Ground truth, prediction, and residuals for Cardiff.
Units are crime incidents per square kilometer.
Table 3.
Mean performance of the baseline alcohol-outlet only linear regression vs the proposed LASSO-based POI model across 100 random test/train splits.
Table 4.
Occurence count of POI classes occurring in top ten model coefficients.
Table 5.
Top 10 LASSO R2 coefficients, combined model.
Table 6.
Combined model performance.