Abstract
Accurately determining responsibility in traffic accidents is crucial for ensuring fairness in law enforcement and optimizing responsibility standards. Traditional methods predominantly rely on subjective judgments, such as eyewitness testimonies and police investigations, which can introduce biases and lack objectivity. To address these limitations, we propose the AMFormer(Arithmetic Feature Interaction Transformer) framework-a deep learning model designed for robust and interpretable traffic accident responsibility prediction. By capturing complex interactions among key factors through spatiotemporal feature modeling, this framework facilitates precise multi-label classification of accident responsibility. Furthermore, we employ SHAP (SHapley Additive Interpretation) analysis to improve transparency by identifying the most influential features in attribution of responsibility, and provide an in-depth analysis of key features and how they combine to significantly influence attribution of responsibility. Experiments conducted on real-world datasets demonstrate that AMFormer outperforms both other deep learning models and traditional approaches, achieving an accuracy of 93.46% and an F1-Score of 93%. This framework not only enhances the credibility of traffic accident responsibility attribution but also establishes a foundation for future research into autonomous vehicle responsibility.