Data scientists warned against deleting features based solely on univariate AUC

Data science experts caution against removing features from models based solely on their individual predictive power. In fields like banking and healthcare, valuable signals often emerge from feature interactions, not isolated metrics. Models like gradient boosting trees can identify critical patterns that univariate analysis, such as AUC, might miss. Deleting features scoring near random noise can therefore harm model performance in complex, real-world datasets.