AdaBoost is a widely-used boosting algorithm known for its effectiveness in improving model accuracy by iteratively adjusting sample weights to focus on misclassified examples. However, its performance deteriorates in the presence of noisy and imbalanced data. To address these challenges, we propose Adaptive Parameter Pair Boosting (APPBoost) to dynamically adjust sample weights by introducing two parameters during the boosting process. APPBoost employs a new loss function with the parameter pair \((\theta , \epsilon )\) , and utilizes an updated weight adjustment mechanism to optimize the classification process in each iteration. This approach mitigates the impact of noisy and misclassified samples, leading to improved generalization and robustness. Theoretical analysis demonstrates that APPBoost effectively bounds the training error and is a special case of the Forward Stagewise Additive Model. Experimental results on both simulated and real-world datasets show that APPBoost outperforms traditional AdaBoost and its variants, particularly in scenarios with high noise levels and class imbalance. In a case study on cardiovascular disease diagnosis, coupled with RFE-RF feature selection, APPBoost achieves superior accuracy, precision, and AUC compared to conventional machine learning and deep learning methods, as well as a 3.2% improvement over AWABoost and a 7.8% improvement over AdaBoost. Additionally, SHAP analysis provides interpretability by quantifying feature contributions, offering insights into the decision-making process of APPBoost.
