Spatiotemporal Urban Crime Prediction Using PSO-Enhanced GCN-CBAM-BiLSTM

Urban crime prediction remains a major challenge due to the complex spatial and temporal dependencies associated with crime events in large urban environments. Existing deep learning models often suffer from overfitting, unstable convergence, high computational complexity, and poor minority class detection. This study proposes a PSO-enhanced Hybrid GCN–CBAM–BiLSTM framework for scalable spatiotemporal crime prediction. The framework integrates Graph Convolutional Networks (GCN) for spatial dependency learning, Convolutional Block Attention Module (CBAM) for adaptive feature refinement, Bidirectional Long Short-Term Memory (BiLSTM) for temporal sequence modeling, and Particle Swarm Optimization (PSO) for intelligent hyperparameter optimization and convergence improvement. Historical urban crime records were preprocessed and represented as graph-structured spatial–temporal data. The proposed model was trained and evaluated using accuracy, precision, recall, F1-score, RMSE, and computational efficiency metrics. Experimental results showed that the optimized framework achieved a test accuracy of 94.03%, test loss of 0.3529, macro-average F1-score of 0.72, and weighted-average F1-score of 0.94. The model also recorded an R² value of 0.7963 and RMSE of 0.1466, indicating improved predictive reliability and reduced error. Furthermore, the framework demonstrated stable convergence, minimal overfitting, reduced training–validation gap, and improved minority class detection compared to existing approaches. The findings confirm that the proposed PSO-enhanced Hybrid GCN–CBAM–BiLSTM framework provides an efficient and scalable solution for intelligent urban crime forecasting and real-time decision-support applications.