Traditional models such as SVM or decision trees are extremely sensitive to label noise, because their capacity and optimization path lack the preference for learning commonalities first and memorizing special cases later.