Graph-Based Clustering and Explainable Graph Neural Networks for Crime Hotspot Prediction in Bangladesh (2020–2025)

Bangladesh has experienced noticeable variation in crime across its metropolitan and range-level policing units during 2020–2025. Some regions, such as Dhaka and Chattogram, regularly report high volumes of violent and property crimes, while others show more sporadic spikes. Conventional analysis tools, for example, trend plots or static heatmaps, often describe these patterns in isolation and struggle to capture how crime in one area resembles or relates to crime in another. This paper presents a graph-based framework that combines community detection, Graph Neural Networks (GNNs), traditional machine learning, and explainable AI (XAI) to anticipate high-risk regions and to understand how policing units are connected in terms of crime behaviour. Monthly crime data from 17 units are used to construct a similarity graph where nodes represent units and edges represent similarity in crime statistics. Louvain and spectral clustering reveal three stable crime-behaviour communities, and the same engineered features feed Random Forest, Graph Convolutional Network (GCN), and Graph Attention Network (GAT) models for next-month high-risk prediction. Random Forest achieves the highest overall accuracy, while the GAT model attains the highest recall on high-risk months, making it useful where missing an emerging hotspot is costlier than a false alarm. SHAP and GNN Explainer are applied to interpret the Random Forest and GCN respectively, showing that Random Forest concentrates its decisions on a small set of temporal features while the GCN distributes importance more evenly across features and relies on a specific subset of high-weight connections between policing units (importance-ranking correlation between the two methods = 0.5125). The results indicate that graph-based analysis, combined with explainability tools, can make crime forecasts more interpretable and useful for planning, even when a classic tabular model remains the most accurate overall predictor.