Predikce kriminality s využitím strojového učení

Abstract

In practice criminality prediction can significantly improve strategic positioning of police patrol in the city, which helps prevent crime from occurring. Machine learning is one of the most widely used method for this problem. However, there is still need to keep comparing various types of algorithms and approaches to get better results. This thesis compares several types of algorithms. Models was learned from data provided by Police of Czech Republic (PČR) for the years 2020 and 2021 on the territory of the city Ostrava. Only selected categories of crimes are entered into the models: theft, burglary, other property crimes and offences against property according to §50. Several methods for resampling the unbalanced dataset were compared in this paper. SMOTETomek was chosen as the best method. It was found that more complex algorithms, such as boosting decision trees or neural networks yield more effective results.