Ada-GCNLSTM: An adaptive urban crime spatiotemporal prediction model

Abstract

Accurate crime prediction is crucial for the proactive allocation of law enforcement resources and ensuring urban safety. A major challenge in achieving accurate predictions lies in identifying generalized patterns of criminal behavior from spatiotemporal features in crime data. Additionally, the inherent randomness and volatility of crime data at the spatiotemporal level introduce noise, which can mislead prediction models. While many effective spatiotemporal crime prediction methods have been proposed, most overlook this issue, reducing their ability to generalize. In this paper, we introduce a novel deep learning-based model, adaptive-GCNLSTM (Ada-GCNLSTM). Specifically, in the spatial feature extraction module, we enhance the model's ability to capture crime spatial distributions by leveraging graph convolutional networks to model spatial dependencies in conjunction with the maximum mean discrepancy to extract the universal features of crime data. We then incorporate a memory network based on long short-term memory network to capture the underlying relationships between temporal features. Through extensive experiments, our model demonstrates an average improvement of 11.7% in mean absolute error and 2.7% in root mean squared error across the three datasets, outperforming the best baseline model. These results underscore the effectiveness of our approach in enhancing crime prediction accuracy.