Spatio-Temporal Graph Neural Networks for Accurate Crime Prediction

As a matter of public safety and resource allocation, crime prediction is of paramount importance. As a result of applying data preprocessing techniques and a graph-based approach, this paper presents a crime prediction model. This study shows a comprehensive analysis of crime incidents reported in Chicago over five years, utilizing several preprocessing steps in preparing the dataset. Irrelevant features are eliminated, missing values are handled, and a new feature is extracted. An algorithm based on the k-nearest neighbors method is proposed for representing crime incidents by constructing a graph representation. Each crime incident is connected to its k nearest neighbors based on spatial coordinates. This study compares graph-based machine learning algorithms and conventional approaches, including logistic regression, decision trees, and the recently related work, Spatial-Temporal Meta-path Guided Explainable Crime Prediction (STMEC). The results reveal that the graph-based model demonstrates superior accuracy, recall, and F1 scores compared to traditional models, notably outperforming the recently related work, STMEC. Based on the current study, the graph-based approach effectively captures crime data's spatial dependencies and patterns. Incorporating additional features and fine-tuning hyperparameters of the model can provide valuable insights into crime prevention strategies and urban planning applications.