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

**Authors:** Isha Hameed, Samuel Sharpe, Daniel Barcklow, Justin Au-Yeung, Sahil Verma, Jocelyn Huang, Brian Barr, C. Bayan Bruss  
**arXiv:** [2207.05566](https://arxiv.org/abs/2207.05566)  
**Date:** 2022-07-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2207.05566v1/

## Key points

- Lack of comprehensive ground truth sources in XAI poses a challenge for determining effective methods and hyperparameters
- Ablation studies are valuable for evaluating XAI methods by perturbing inputs to assess model sensitivity
- Researchers focus on applying ablation studies to models on various tabular datasets with different feature types
- Differentiable models are utilized to expand the scope of XAI methods, but principles can also be applied to non-differentiable models
- Contributions include distinguishing between ablation perturbations and XAI baseline distributions, treating categorical features in label encoded form, and proposing guardrails for ablation studies
- A more rigorous approach is offered for conducting ablation studies on tabular data, raising important questions for future research in explainable artificial intelligence

## Summary

In the realm of explainable artificial intelligence (XAI), the lack of comprehensive ground truth sources has posed a challenge in determining the most effective methods and hyperparameters for specific use cases. Ablation studies have emerged as a valuable tool to assess the efficacy of XAI methods by evaluating the sensitivity of model performance when inputs are perturbed. The intuitive premise is that perturbing important features identified by an XAI method should result in larger decreases in model capability compared to less important features. To address this issue, researchers have focused on applying ablation studies to models built on various tabular datasets with different feature types. By utilizing differentiable models, they aim to expand the scope of XAI methods available for experimentation. However, these principles can also be applied to non-differentiable models. The contributions of this research include distinguishing between ablation perturbations and XAI baseline distributions to eliminate confounding effects, emphasizing the importance of treating categorical features in their label encoded form during perturbations for accurate feature ranking, and proposing guardrails or sanity checks for ablation studies to define a feasible region and draw more robust conclusions. Overall, these contributions offer a more rigorous approach to conducting ablation studies on tabular data and raise important questions for future research in the field of explainable artificial intelligence. The study sheds light on the complexities involved in assessing XAI methods' performance and highlights the need for standardized frameworks for application and validation across different methodologies within post-hoc explainability efforts.