AUTHOR=Qin Zhenkai , Wei Baozhong , Gao Caifeng , Zhu Feng , Qin Weiqi , Zhang Qian 

TITLE=ACSAformer: A crime forecasting model based on sparse attention and adaptive graph convolution

JOURNAL=Frontiers in Physics

VOLUME=Volume 13 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/physics/articles/10.3389/fphy.2025.1596987

DOI=10.3389/fphy.2025.1596987

ISSN=2296-424X

ABSTRACT=IntroductionCrime forecasting is crucial for urban safety management, as it facilitates the optimization of police resource allocation, crime prevention, and the enhancement of public security. However, existing supervised learning methods encounter several limitations in processing crime data, including inadequate spatiotemporal representation capabilities, poor generalization and robustness, and high computational complexity, all of which hinder forecasting efficiency.MethodsTo address these challenges, this paper proposes a deep learning-based spatiotemporal sequence forecasting model, named ACSAformer. The model integrates the Transformer architecture with adaptive graph convolutional layers and a sparse attention mechanism. The incorporation of adaptive graph convolution significantly enhances the model’s ability to represent multivariate spatiotemporal sequences, enabling it to capture complex inter-feature relationships and dynamic correlations, thereby improving generalization and predictive accuracy. The sparse attention mechanism further reduces the number of key tokens each query needs to attend to by computing similarity scores only for query-key pairs selected according to predefined patterns, reducing the computational complexity from O(L2) to O(L log L) and greatly improving the efficiency of long-sequence processing. Extensive experiments were conducted on five real-world crime datasets—four from Los Angeles and one from Chicago—covering the period from 2020 to 2023.ResultsThe results demonstrate the superior performance of ACSAformer compared to traditional spatiotemporal forecasting models across multiple evaluation metrics. Specifically, on the DS1 dataset, the proposed model achieved a 17.6% reduction in Mean Squared Error (MSE) and a 9.2% reduction in Mean Absolute Error (MAE).DiscussionThese findings confirm that ACSAformer not only improves predictive accuracy and robustness but also offers better computational efficiency, showcasing its potential for application in complex spatiotemporal tasks such as crime forecasting.