Minimal Data, Maximum Clarity: A Heuristic for Explaining Optimization

Less data can mean better explanations: EZR demonstrates that using fewer but more carefully selected data points produces optimization models that are both effective and easier to understand, achieving over 90% of best-known performance while requiring significantly fewer labeled examples than traditional approaches.

Active learning reduces labeling burden: By strategically identifying which configurations are most informative to label (using Naive Bayes sampling), EZR substantially reduces the costly and error-prone process of manual labeling that practitioners typically face when optimizing complex software systems.

Decision trees outperform black-box explanations: The paper shows that EZR's transparent decision tree explanations provide clearer, more actionable insights than standard explainable AI methods like LIME and SHAP, offering "cohort-based rationales" that help practitioners understand why certain configurations are recommended.

Practical efficiency for real-world systems: Tested across 60 real-world datasets, EZR provides a lightweight, modular framework that balances performance with interpretability—addressing a critical gap where practitioners must navigate vast configuration spaces without sacrificing their ability to understand and trust optimization decisions.

Transparency supports responsible deployment: By making optimization logic explicit and understandable through clear explanations, EZR enables practitioners to audit decisions, identify potential biases, and maintain meaningful human oversight over automated configuration choices in software systems.