xAI-Drop: Don't Use What You Cannot Explain

Explainability can improve AI model performance: The research demonstrates that using explainability methods to identify which parts of a neural network are actually contributing useful information—rather than just guessing—leads to better accuracy and more reliable predictions on real-world tasks.

Current AI systems often learn from "noise": Graph Neural Networks (GNNs), which are used for analyzing interconnected data like social networks and biological systems, tend to pick up on irrelevant patterns and unnecessary complexity, similar to how a student might memorize trivial details instead of core concepts. The xAI-Drop method addresses this by removing these unhelpful elements.

Explainability should be built into training, not added later: Rather than treating explainability as an afterthought to audit a finished model, this work argues that understanding why a model makes decisions should guide how it learns from the beginning, making the system more trustworthy throughout development.

Better explanations lead to better generalization: The paper shows that models trained with explainability-guided filtering not only perform better on new, unseen data but also produce higher-quality explanations of their decisions—addressing two critical concerns for deploying AI in high-stakes applications like healthcare or finance.