Personalized prediction of unsafe driving behaviors for drivers of dangerous goods transportation trucks based on an attribute graph interaction model.

With the growing deployment of dangerous goods transportation trucks (DGTTs), ensuring driving safety has become increasingly important. Given the high disaster potential and hazardous nature of DGTTs, source-level risk control is essential. To support proactive risk management at the source, we propose a method for predicting unsafe driving behaviors before trips. This method leverages trajectory data and intelligent video collected from legally mandated on-board terminals. We adopt a recommender system (RS) approach for its capacity to capture intricate attribute interactions and provide personalized predictions. Drawing an analogy between RS components and our scenario, drivers correspond to users and alarms to items, with their respective attributes forming two sides of the model. We introduce a Bilateral Graph Interaction-based Collaborative Filtering (BGICF) model, enhanced with Adversarial Graph Dropout (AdvDrop). BGICF models both internal coupling and external interaction between attributes. Furthermore, to address attribute popularity bias and improve interpretability in BGICF, we integrate AdvDrop, which constructs bias-mitigating and bias-aware subgraphs using a bias measurement function and optimizes them through adversarial learning. We collected natural driving data from an active safety platform from 23 DGTT companies in Beijing, China, covering over 58 million trajectory points and 211,157 alarm records. Experimental results showed that BGICF-AdvDrop achieves macro precision, recall, F1-score, and accuracy of 0.8202, 0.8114, 0.8101, and 0.8416, respectively, outperforming other models while providing better interpretability.