GCN-GRU: A Spatiotemporal Deep Learning Framework Integrating Graph Convolution and Gated Recurrent Networks for Crime Prediction

Law enforcement agencies rely on accurate crime prediction systems to study past and present crime trends in order to forecast and prevent future incidents. Among Deep Learning (DL) approaches, time series prediction using Long Short-Term Memory (LSTM) networks is popular because modeling long-term temporal dependencies and sequential patterns is necessary for crime data. However, LSTM struggles with large number of parameters due to three gates, difficulty in capturing very short-term dependencies and increased memory consumption, limits the prediction on real-time crime datasets. For spatial learning, Graph Convolutional Networks (GCNs) have been used to capture crime area based correlations and spatial dependencies in crime data. However, GCN often overfit to local graph structures, struggle to extract transferable features across diverse regions and exhibit reduced performance when spatial data is noisy or incomplete. To overcome such limitations a Graph Convolutional Network with Gated Recurrent Unit (GCN-GRU) is put forward in this paper to enhance crime prediction. In this model, GCN dynamically adapts the graph topology based on spatial data characteristics to extract relevant features across diverse spatial regions in the crime dataset. Also, this mechanism captures both local and global spatial dependencies improve resilient to noisy or incomplete data. By updating neighborhood relationships during training, GCN avoids dependence on fixed local structures reducing overfitting and improving spatial feature stability. GRU employs only two gates (reset and update) with fewer parameters enabling faster training and lower memory usage. Moreover, the reset gate enhances the handling of sudden and short-term variations in sequential crime data while preserving the ability to technique long-standing needs. In the temporal modeling module, GRU network captures the underlying relationships between sequential crime events and their temporal patterns. Along with this Cross-Entropy Loss function is employed to help the method to give greater probabilities to correct crime categories to improve classification accuracy and enhance decision confidence in crime prediction. Thus, GCN improves spatial feature mapping and GRU enhances temporal sequence learning in enhanced crime classification. Experimental results demonstrate that the proposed GCN-GRU outperforms existing baseline approaches in crime prediction.