X-SHIELD: Regularization for eXplainable Artificial Intelligence

Explainability can improve model performance, not just transparency: X-SHIELD demonstrates that using explanations to guide model training—rather than only explaining finished models—can simultaneously boost both a model's accuracy and its interpretability, challenging the assumption that these goals are separate concerns.

The method works by strategically hiding features during training: X-SHIELD forces AI models to learn more robust patterns by deliberately hiding certain input features during the training process, which helps models generalize better and rely less on potentially unreliable or biased features.

Explanations can be used as actionable feedback for training: Rather than treating explanations as a post-hoc analysis tool, this work shows that explanation methods can be integrated directly into the model-building process to guide which features the model should learn to depend on, creating a feedback loop that improves model quality.

The approach is practical and integrates into existing workflows: X-SHIELD functions as a regularization technique that can be added to standard model training objectives without requiring major architectural changes, making it potentially adoptable across different AI applications and domains.

Experimental validation confirms dual benefits: Testing on benchmark datasets shows that models trained with X-SHIELD regularization outperform standard models on both performance metrics and explainability measures, suggesting the approach delivers on its promise of improving model quality while maintaining transparency.