Towards spatio-temporal crime events prediction

Abstract

The importance of early prediction in reducing the impact of crime cannot be overstated. Machine learning algorithms have proven to be effective in this regard, but their inability to capture key features automatically can be a hindrance. To overcome this challenge, we propose a deep neural network model that is capable of extracting salient features automatically for predicting crime categories using real-world crime data sourced from the Chicago open data portal. To ensure the robustness of our proposed model, we carried out an extensive exploratory data analysis to determine the impact of socioeconomic indicators on crime occurrences. Additionally, we implemented a data upsampling technique to handle class imbalance issues, and we leveraged hyperparameter optimization algorithms to fine-tune the model. The results of our study were impressive. Our proposed model outperformed the baseline model and other algorithms, with an average improvement of 6% in macro F1 score. This suggests that our model is highly effective, if not superior, in predicting crime categories accurately. Overall, our study provides a solid framework for using deep neural network models in crime prediction, while highlighting the importance of automatic feature extraction in enhancing the accuracy of predictions. By reducing the impact of crime through early prediction, we can help to create a safer and more secure society.