STCAGNN-RNKDE: a traffic accident prediction model for spatiotemporal combinatorial attention graph neural networks using Ripley’s K and network kernel density estimation

Existing research on road traffic accident prediction faces challenges such as unreasonable prediction target selection, incomplete integration of spatiotemporal information from multi†source risk factors, inadequate consideration of factor sparsity, and insufficient characterization of spatial spillover effects on road risk. To address these issues, this paper proposes a spatiotemporal combinatorial attention graph neural network that integrates Network Ripley’s K function and Network Kernel Density Estimation (STCAGNN-RNKDE). By constructing road network subgraphs, the model achieves fine-grained temporal accident prediction at the actual road network level and incorporates three risk exposure features, traffic flow, traffic violations, and real-time travel intensity, into its predictive dimensions. Dedicated extraction and fusion modules then process heterogeneous risk factors across spatial and temporal dimensions. Moreover, Ripley’s K-NKDE is applied during data preprocessing to tackle the sparsity of influencing factors and accidents while capturing their spatial spillover effects for road risk. Experimental results indicate that 5-hop neighbor aggregation produces the optimal subgraph. Overall, our model outperforms baselines, achieving a 7.4 % recall gain, and it also demonstrates strong temporal robustness. The model structure is complete and reasonable, and it is found that spatiotemporal information > spatial information > temporal information. Ripley’s K-NKDE effectively addresses the sparsity of influencing factors and accidents and characterizes their spatial spillover effects on road traffic risk. Among risk exposure features, traffic violations > real-time travel intensity > traffic flow, and regarding overall risk information, spatiotemporal features > spatial features > temporal features.