START: A Spatiotemporal Autoregressive Transformer for Enhancing Crime Prediction Accuracy
Umair Muneer Butt, Sukumar Letchmunan, Mubashir Ali, Hafiz Husnain Raza Sherazi · IEEE Transactions on Computational Social Systems · 2025
This study presents a spatiotemporal autoregressive transformer (START) that combines the strengths of spatiotemporal transformers and vector autoregression (VAR) to predict crime by type, handling nonstationary data and capturing long-term dependencies. Moreover, Start utilizes one-hot encoding to differentiate crime types, enabling the model to assign specific attention to individual crime categories. Seasonal trend decomposition using LOESS (STL) is employed to decompose time series into trend, seasonality, and remainder components by effectively capturing residual variations. VAR attention enhances prediction accuracy by focusing on specific crime types and prioritizing relevant information. The proposed approach is evaluated on publicly available crime datasets from New York, Chicago, and Los Angeles and three state-of-the-art evaluation measures: mean error (ME), mean absolute percentage error (MAPE), and root mean square error (RMSE). Our simulation results demonstrate the significance of the proposed method in enhancing crime prediction accuracy for the most critical crimes compared to state-of-the-art methods. Enhanced crime prediction may help law enforcement agencies allocate resources efficiently by focusing on the most critical areas.