An adaptive data rebalancing framework for real-time traffic risk prediction

This study proposes a framework to mitigate data imbalance in risk prediction, providing an adaptive data rebalancing approach. Specifically, the Synthetic Minority Over-sampling Technique (SMOTE) is applied for oversampling, and Random Undersampling (RU) for undersampling, using various fixed balancing ratios. Subsequently, the original and rebalanced data are input into four models to evaluate the impact of data rebalancing on prediction. Additionally, a Genetic Algorithm (GA) is introduced to identify the optimal ratio, with the explicit objective of jointly optimizing prediction precision and efficiency to achieve the best overall performance. Ultimately, convergence curve analysis and robustness checks are conducted to validate the stability and reliability of the results. Results show that data rebalancing enhances prediction performance, particularly after SMOTE processing, with a 2:1 ratio yielding the best outcomes. After GA optimization, the Gated Recurrent Unit (GRU) model consistently performs the best among the models processed with SMOTE and RU, with optimal ratios identified as 2.3:1 and 2.7:1. Finally, the reliability of the optimization is ensured by analyzing the convergence curves, which demonstrate a stable decrease in fitness values over iterations, thereby mitigating the risk of local optima. Additionally, robustness analysis validates the stability of the results under minor fluctuations in proportions.