Spatio-temporal police incident analysis with explainable AI

Sammanfattning

Police incidents and the public safety are remaining as essential concerns that are societal, especially within urban environments where there are patterns of incidents that vary through location and time. In conjunction with the availability of data within the public sector together with progresses within artificial intelligence (AI), new opportunities have been created for conducting analysis of patterns of police incidents that are spatio-temporal. The aim of the thesis was investigating the extent to which models of machine learning and deep learning approaches can predict the counts of police incidents through diverse locations and time periods within Sweden. Additionally, the thesis explored which factors that were temporal and spatial that were contributing the most to the predictions of the models with explainable artificial intelligence (XAI). Historical data of police incidents was retrieved out of the public API that was provided by the Swedish Police Authority (Polismyndigheten). Decision tree, random forest, XGBoost, long short-term memory (LSTM) and convolutional neural network (CNN) were the models that were implemented and also evaluated by utilizing metrics for regression including R2-score, MAE and RMSE. An explainability analysis was applied with the aim of improving the models’ interpretability. The results indicated that the approaches of traditional machine learning achieved a stronger performance of predictions compared to the approaches of deep learning. XGBoost achieved the overall strongest performance, followed by random forest being close, while LSTM and CNN achieved a performance that was weaker for the available data. The analysis of explainability showed that both of the variables that were temporal and spatial had an influence on the predictions where the factors that were geographical contributed the strongest. The findings in the thesis suggest that techniques of machine learning, especially models that are tree based are able to provide insights that are valuable to patterns of police incidents that are spatio-temporal. Furthermore, the thesis illustrated how XAI have the possibility of improving the transparency and also the understanding of the models in contexts of the public sector.