For Those Who Believe in Faithfulness: Optimizing the Area Under Insertion and Deletion Curves for Ranking Relative Feature Importance

Attribution methods help explain AI decisions through "heat maps" that show which input features matter most for a prediction. This paper addresses a critical need in responsible AI: ensuring these explanations are actually faithful to how the model works, rather than just plausible-sounding stories.

The paper proposes Ra-NEM, a new method that measures explanation quality by testing what happens when features are systematically added to or removed from inputs. This "insertion and deletion" approach provides a concrete, measurable way to verify whether highlighted features truly drive the model's predictions—essential for building trust in high-stakes applications.

Ra-NEM can be applied to any machine learning model without changing how the model performs, making it practical for real-world deployment. The method also runs quickly enough for online applications, addressing a common barrier to adopting explainability tools in production systems.

The research demonstrates that Ra-NEM produces more faithful explanations than competing methods while maintaining strong performance on other explainability metrics. This suggests the approach offers a genuine improvement rather than trading off one desirable property for another.

For socially relevant tasks (healthcare, criminal justice, lending, etc.), having reliable explanations is foundational to responsible AI. This work contributes a technically rigorous tool for practitioners to verify their models' explanations are trustworthy before deployment.