An interpretable dynamic ensemble selection multiclass imbalance approach with ensemble imbalance learning for predicting road crash injury severity.

Accurate prediction of crash injury severity and understanding the seriousness of multi-classification injuries is vital for informing authorities and the public. This Knowledge is crucial for enhancing road safety and reducing congestion, as different levels of injury necessitate distinct interventions, policies and responses to support sustainable transportation. Existing ML techniques often face class imbalance issues, resulting in suboptimal performance. Multi-class imbalance, more challenging than two-class imbalance, is frequently overlooked in traffic risk assessments. To accurately estimate the multi-class accident injuries and comprehend their severity we proposed a novel method called Bayesian Optimized Dynamic Ensemble Selection for Multi-Class Imbalance (DES-MI) with Ensemble Imbalance Learning (EIL), which involves; generating a pool of base classifiers with EIL methods and utilizing DES-MI to choose suitable classifiers. Utilizing homogeneous and heterogeneous pools of EIL classifiers, our findings demonstrate that DES-MI with EIL considerably enhances classification performance for datasets with multi-class imbalances. DES-MI with Heterogeneous EIL outperformed in performances while DES-MI with BRF achieved notable results in homogeneous ensembles. Important variables including road user gender, occupant age, month, airbag deployment, and road profile are also identified using SHAP for interpretability. Our DES-MI model with EIL classifiers and SHAP, by addressing multi-class imbalance, offers insightful information to stakeholders in road traffic safety by supporting the development of safer, efficient and sustainable urban road transport systems.