Augmented graph information bottleneck with type-aware periodicity heterogeneity for explainable crime prediction

Accurate crime prediction is crucial for proactive law enforcement, enhancing public safety, and optimizing resource allocation. However, existing solutions struggle with two key challenges: (1) crime periodicity heterogeneity, where different types of crimes exhibit distinct periodic patterns; and (2) lacking of post-hoc explainability, as the black-box effect of deep neural models makes it challenging to interpret the model performance. To respond to these concerns, we present EX-Crime, a novel solution for crime prediction with an inherently explainable view. By following the merits of information bottleneck (IB) principles, we introduce an augmented graph information bottleneck learning block that is coupled with an integrated crime pattern modeling to capture spatial–temporal patterns of crimes. Moreover, a new crime type-aware periodicity learning, employing Fast Fourier Transform and autocorrelation techniques, is proposed to identify the periodic patterns of distinct crime for enhanced temporal modeling. Our extensive experiments on real-world datasets demonstrate that EX-Crime gains substantial improvements in prediction accuracy and interpretability, uncovering its potential to significantly advance crime understanding across diverse urban settings.