Variable selection (VS) is a critical step in developing predictive binary classification (BC) models. Many traditional methods for assessing the added value of a candidate variable provide global performance summaries and lack an interpretable graphical summary of results. To address this limitation, we developed the U-smile method, a residual-based, post hoc evaluation approach for assessing prediction improvements and worsening separately for events and non-events. The U-smile method produces three families of interpretable BA-RB-I coefficients at three levels of generality and a standardized graphical summary through U-smile and prediction improvement–worsening (PIW) plots, enabling transparent, interpretable, and explainable VS. Validated in balanced and imbalanced BC scenarios, the method proved robust to class imbalance and collinearity, and more sensitive than traditional metrics in detecting subtle but meaningful effects. Moreover, the method’s intuitive visual output (U-smile plot) facilitates the rapid communication of results to non-technical stakeholders, bridging the gap between data science and applied decision-making. The U-smile method supports both local and global evaluations and complements existing explainable machine learning (XML) and artificial intelligence (XAI) tools without overlapping in their functions. The U-smile method offers a transparency-enhancing and human-oriented approach for ethical and fair VS, making it highly suited for high-stakes domains, e.g., healthcare and public health.
Kubiak et al. (Fri,) studied this question.