Modified resampling strategy for extreme values in imbalanced air pollution data using moving block bootstrapping approach with relevance weighting (MBB-RW)

Air pollution datasets typically exhibit a right-skewed distribution. These conditions are caused by the presence of extreme events leading to imbalanced data distribution. This imbalanced regression poses a notable challenge in predictive modeling, as the models tend to be biased towards frequent normal events while underperforming extreme events. Therefore, to address this issue, the resampling approach is crucial in handling these extreme events to improve model performance. In this study, a modified resampling strategy, called Moving Block Bootstrapping with Relevance Weighting (MBB-RW), is proposed to address imbalanced regression problems. By integrating MBB with relevance weighting, the time-series dependence of air pollution data is preserved while placing greater emphasis on extreme events. The findings demonstrate that MBB-RW can mitigate data imbalance and enhance model prediction accuracy for extreme events. These enhancements are evident in the performance of the Extreme Gradient Boosting (XGBoost) model in predicting PM₁₀, where the evaluation metrics showed significant reductions after applying MBB-RW: RMSE dropped from 108.3010 to 39.1846 (63.8188%) and MAE from 85.1041 to 27.1082 (68.14700%). The key contribution of this study is the development of MBB-RW resampling strategy designed to improve extreme values in an imbalanced regression dataset, with our focus on the air pollution dataset. Simultaneously, this method can be implemented to enhance the accuracy of PM₁₀ concentration predictions, specifically during extreme air pollution events.