Selecting the most important variables for crime prediction based on machine learning approaches

Crime prediction has been challenging, partly due to the difficulty in selecting predictive variables. Previous studies theoretically and empirically confirmed the relationship between many variables and crime. This study identified a total of 19 theory-driven variables and applied machine learning tree models for crime prediction. To refine the variable pool, we extracted the subsets of the variables from the variable pool by using feature selection processes, which was based on Gain and Permutated Variable Importance Measure (PVIM) methods. We finally found historical cases, percent nonlocal population, and ambient population (25-44) variables ranked high. Thus, our results call more attentions on these variables for crime prediction.