Explainable machine learning models often rely on mixed-integer optimization to enforce explicit sparsity constraints, balancing interpretability against prediction accuracy. However, finding the right sparsity level typically requires computationally expensive procedures such as cross-validation, which limits their practical use.
The paper extends distributionally robust optimization (DRO) regularization—previously developed for convex training formulations—to mixed-integer sparse programs. This extension provides statistical guarantees that depend on the sparsity parameter, allowing strong generalization without the heavy computational burden of traditional tuning methods.
The authors illustrate the approach on explainable binary classification using sets of feature-value rules. This demonstrates a practical path toward building interpretable models that maintain robust out-of-sample performance while avoiding costly model-selection steps.