Enhancing injury severity prediction in small sample scenarios: A novel TabNet-DAAN approach for transfer learning

This study proposes a novel transfer learning framework, TabNet-Dynamic Adversarial Adaptation Networks (TabNet-DAAN), to address the challenge of inadequate and imbalanced traffic crash data in certain regions. By integrating TabNet’s robust feature extraction with the dynamic adversarial framework of DAAN, this approach enhances transfer learning effectiveness through automatic alignment of data distributions across different domains. Moreover, it integrates a Class-balanced Loss (CBLoss) algorithm to mitigate the adverse effects of data imbalance and further enhance the model’s robustness. The methodology was empirically validated using two-vehicle crash data (2018–2022) from Birmingham (7,432 crashes, source domain) and Leeds (3,275 crashes, target domain) in the UK. Results from ablation studies confirm the effectiveness of the CBLoss algorithm and TabNet module. Moreover, TabNet-DAAN demonstrates superior performance compared to other transfer and non-transfer learning algorithms in predicting injury severity of target domain data, exhibiting excellent generalization capabilities, particularly when using a 7:1 source-to-target data ratio in training. Additionally, feature importance analysis shows that TabNet-DAAN adaptively refines its feature selection, effectively identifying key predictors across domains. These findings highlight the model’s potential to improve predictions of injury severity in data-scarce regions, presenting profound implications for targeted traffic safety interventions.