Robust-GBDT: leveraging robust loss for noisy and imbalanced classification with GBDT

Label noise and class imbalance are persistent challenges in tabular classification tasks. We propose Robust-GBDT, a Gradient Boosting Decision Tree framework that integrates nonconvex loss functions to improve robustness under such adverse conditions. We establish theoretical conditions under which nonconvex losses can be safely incorporated into GBDT models, enabling reliable optimization despite nonconvexity. Robust-GBDT extends existing GBDT methods by addressing limitations in prior robust boosting techniques, including support for multi-class classification, handling imbalanced data, and managing missing values. Empirical results demonstrate that Robust-GBDT consistently outperforms state-of-the-art bagging and robust boosting approaches, with average gains of 1.33% in binary and 0.51% in multi-class classification, and maximum improvements of 8.15% and 10.72% under label noise and class imbalance, respectively. The code is available at https://github.com/Luojiaqimath/Robust-GBDT.

Duke Scholars

Author Shixin Xu DKU Faculty