Enhancing Feature Selection and Interpretability in AI Regression Tasks Through Feature Attribution

Feature attribution methods can improve AI performance, not just explain it: The paper demonstrates that explainability techniques like Integrated Gradients can be used proactively to filter out unhelpful data features, leading to more accurate and stable predictions in regression tasks—showing that transparency and performance improvements go hand-in-hand.

XAI research has overlooked regression problems: Most existing explainability work focuses on classification (like image recognition or fraud detection), but this study highlights that regression tasks—which predict continuous values like temperatures or vibration levels—deserve equal attention for interpretability and deserve equal attention for interpretability.

A practical pipeline for real-world applications: The researchers developed a concrete, implementable approach combining Integrated Gradients with k-means clustering that was tested on actual industrial data (turbomachinery blade vibration), suggesting these methods can move beyond academic exercises into practical engineering applications.

Robustness matters alongside accuracy: By removing uninformative features, the approach improves not just prediction accuracy but also stability—meaning AI systems become more reliable and less prone to erratic behavior when faced with new or slightly different data, which is critical for safety-sensitive industrial applications.