Interpretable multi-variable transformer network for regional-level short-term bicycle crash risk prediction

Abstract

Effective short-term prediction of bicycle crashes at the urban regional level is critical for proactive infrastructure safety interventions and data-driven traffic management. However, three key challenges persist: (1) inadequate modeling of complex spatiotemporal dependencies in multi-source heterogeneous data; (2) poor handling of extreme class imbalance and lack of interpretability in deep learning-based short-term predictions; and (3) limited exploration of bicycle infrastructure’s role in regional crash risk assessment. In response to these challenges, we propose an Interpretable Multi-variable Transformer Network (IMTN) that employs four specialized Transformer encoder blocks to extract spatial and temporal dependencies from heterogeneous inputs. To mitigate the severe class imbalance, our approach uses a single, shared model to predict crash risk for one region at a time, rather than all regions simultaneously. This reformulation avoids data sparsity while retaining multi-region inputs, and a spatial weighting mechanism is used to preserve inter-regional dependencies. Meanwhile, an improved Layer-wise Relevance Propagation (LRP) framework is employed to enhance the interpretability of IMTN. We conduct our experiments on a four-year dataset from London, which includes crash records, public bicycle trips, time, weather, road networks, land use, and a rich set of 48 bicycle infrastructure features. The model comparison demonstrates that IMTN consistently outperforms competitive baselines, reducing false positive rate (FPR) by up to 9.08%, improving the area under the curve (AUC) by up to 3.49%, and increasing the G-mean by up to 5.39%. Our model achieves the best performance at the finest temporal resolution (1-hour aggregation), contrary to common expectations. This suggests that the proposed class imbalance handling method may enhance model performance in high-resolution settings. In addition, interpretability analysis identifies segregated cycle lanes, Sheffield stands, and colored path markings as high-impact infrastructure variables, providing data-driven insights that can help inform urban safety planning.