A Multidimensional Framework for Traffic Accident Conse-quence Prediction: Integrating Multi-objective Optimization, Explainable AI, and Causal Inference

The original dataset consists of road traffic ac-cident records, road environment and traffic facility attributes, and grid-scale point-of-interest (POI) data, which were spatially matched using accident location co-ordinates, road codes, and grid IDs. After data cleaning, a total of 4,237 valid observa-tions were obtained.

The dataset was randomly divided into training and testing sets at a ratio of 80% and 20%, with the random seed set to 42. The final training set contained 3,389 observations, while the testing set contained 848 observations. Hyperparameter search and Voting ensemble weight optimization were conducted only within the training set, using three-fold cross-validation with random shuffling.