Better Model Selection with a new Definition of Feature Importance

Feature importance metrics matter for responsible AI: The paper demonstrates that how we measure which features drive model predictions directly impacts model selection quality, highlighting that technical choices in explainability tools have real consequences for which models get deployed.

New approach improves both speed and accuracy: The proposed method using "Coefficient of Variation of feature weight" outperforms standard cross-validation in model selection, suggesting that better measurement techniques can make AI systems both more efficient and more reliable without requiring additional computational resources.

Consistency across predictions is a key indicator: By measuring how a feature's importance varies across different samples (rather than just its average importance), the method captures whether a model relies on features in a stable, predictable way—an important signal for trustworthiness that traditional approaches may miss.

Better model selection supports responsible deployment: Since feature importance is explicitly used in model selection decisions, improving this measurement helps ensure organizations choose models that are more explainable and interpretable, supporting the goal of more transparent AI systems.

Practical benefits for XAI workflows: The approach offers time efficiency gains alongside accuracy improvements, making it more feasible for organizations to conduct thorough model evaluation before deployment—a key step in responsible AI governance.