ITBoost: Information-Theoretic Trust for Robust Boosting
Researchers have introduced ITBoost, a novel approach to gradient boosting designed to enhance robustness against noisy labels in tabular data. Unlike traditional methods that emphasize samples with large gradients, ITBoost evaluates sample reliability by examining the evolution of residuals across training iterations. By applying the Minimum Description Length principle, ITBoost down-weights samples with irregular residual patterns, treating them as less trustworthy. This method theoretically offers a tighter generalization bound under label noise and empirically demonstrates improved performance on noisy benchmarks while maintaining strong results on clean data. AI
IMPACT Improves robustness of gradient boosting models against noisy labels, potentially enhancing performance in real-world datasets with imperfect labeling.