Fighting biases with dynamic boosting

Résumé : While gradient boosting algorithms are the workhorse of modern industrial machine learning and data science, all current implementations are susceptible to a non-trivial but damaging form of label leakage. It results in a systematic bias in pointwise gradient estimates that lead to reduced accuracy. This paper formally analyzes the issue and presents solutions that produce unbiased pointwise gradient estimates. Experimental results demonstrate that our open-source implementation of gradient boosting that incorporates the proposed algorithm produces state-of-the-art results outperforming popular gradient boosting implementations.