BASED-XAI: Breaking Ablation Studies Down for Explainable Artificial Intelligence

Ablation studies—a common method for testing XAI tools—lack standardized implementation details, meaning researchers may reach different conclusions about the same explanation method depending on how they conduct the test. The paper demonstrates this by showing how varying perturbation strategies and adding simple safeguards can prevent "potentially flawed conclusions."

How categorical variables are handled matters significantly for both explaining AI decisions and validating those explanations. The researchers identify categorical variable treatment as an important consideration that has not received adequate attention in ablation study design for tabular data.

There is no single "ground truth" for evaluating XAI methods, so the field has relied on theoretical principles (axioms) instead. However, the paper emphasizes that for high-stakes applications requiring explainability, relying on theory alone is insufficient—actual validation through rigorous testing procedures is essential.

Choosing the right baseline and perturbation method is critical for ablation studies to work properly. The research provides guidance on identifying useful baselines and viable perturbation strategies, suggesting that not all approaches are equally valid for testing explanation methods.

Implementation details in XAI validation can fail to match their intended purpose, even when the underlying theory is sound. This gap between ideal design and actual practice underscores why careful, standardized procedures for validating explainability methods are necessary before deploying AI in high-stakes domains.