A method for predicting the severity of traffic accidents based on multi-dimensional audit screening

The present invention discloses a traffic accident severity prediction method based on multi-dimensional audit screening, which belongs to the field of traffic data prediction technology. It comprises the following steps: Firstly, through nested cross-validation, the optimal classifier based on the training set is selected as the final prediction model; then, candidate synthetic samples for the minority class are generated, and an integrated evaluation model is constructed using a heterogeneous classification model trained only based on the real samples; next, the audit index values of each candidate sample in four dimensions - label confidence, boundary margin, model consistency, and local density rationality - are calculated; subsequently, normalization is performed within the same category, and the Pareto non-dominated sorting is used to retain the synthetic samples consistent with the preset target quantity while eliminating the rest; finally, the retained synthetic samples are combined with the real samples to train the final prediction model for accident severity prediction. This invention maintains the overall prediction accuracy basically stable while enhancing the model’s recognition ability and prediction robustness for minority accident samples.