Crimes Prediction Using Spatio-Temporal Data and Kernel Density Estimation

Abstract

This study presents a method to predict crimes by using multiple data sources i.e. spatio-temporal crime dataset and zoning district dataset. The contribution of this study lies in the use of Kernel Density Estimation (KDE) and zoning district dataset to address the issue of crimes prediction. The experiments were performed by training Gradient Boosting Machine (GBM) as a classifier on some subset of features. The best result was achieved by using all features including KDE with smoothing and zoning district feature, namely with multiclass logarithmic loss 2.356104 on validation set and 2.35443 on test set.