# XAI-TRIS: Non-linear benchmarks to quantify ML explanation performance

**Authors:** Benedict Clark, Rick Wilming, Stefan Haufe  
**arXiv:** [2306.12816](https://arxiv.org/abs/2306.12816)  
**Date:** 2023-06-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.12816v1/

## Key points

- Growing interest in explainable artificial intelligence (XAI)
- Lack of formal underpinning in XAI methods raises concerns about reliability and interpretability
- Challenge of suppressor variables influencing model predictions
- Need for rigorous evaluation metrics and benchmark datasets capturing realistic correlations between important and unimportant features
- Introduction of benchmark datasets for non-linear classification scenarios with known ground truth explanations
- Definition of quantitative metrics for evaluating explanation performance with few important features and suppressor variables
- Evaluation of sixteen popular XAI methods across three machine learning architectures reveals struggles to outperform random baselines and edge detection techniques
- Contribution to advancing understanding of XAI methods through a comprehensive evaluation framework addressing linear and non-linear settings, as well as the influence of suppressor variables

## Summary

In recent years, there has been a growing interest in the field of explainable artificial intelligence (XAI), which aims to make the decisions of complex machine learning models more understandable to humans. However, the lack of formal underpinning in XAI methods has raised concerns about the reliability and interpretability of these explanations. To address this issue, researchers have started to focus on empirically validating XAI methods using ground truth data. One key challenge in evaluating XAI methods is the presence of suppressor variables, which are features that do not have a direct statistical association with the prediction target but can still influence model predictions by removing noise. Previous studies have shown that popular XAI methods may incorrectly attribute importance to suppressor variables, even in linear settings. This highlights the need for rigorous evaluation metrics and benchmark datasets that capture realistic correlations between important and unimportant features. To address this gap, this study introduces benchmark datasets for three different non-linear classification scenarios where ground truth explanations are explicitly known. These scenarios involve distinguishing between different types of tetrominoes overlaid on noisy backgrounds, requiring non-linear predictive models for effective classification. The study also defines quantitative metrics for evaluating explanation performance in cases with few important features and investigates the impact of suppressor variables introduced through different types of background noise. Furthermore, the study evaluates sixteen popular model-agnostic and model-specific XAI methods across three machine learning architectures to assess their explanation performance. The findings suggest that existing XAI methods often struggle to outperform random baselines and edge detection techniques, highlighting the challenges in interpreting explanations derived from different model architectures. Overall, this research contributes to advancing our understanding of XAI methods by providing a comprehensive evaluation framework that considers both linear and non-linear settings while addressing the influence of suppressor variables on explanation performance. By introducing realistic correlations between important and unimportant features in benchmark datasets, this study sheds light on the limitations and potential improvements needed in current XAI methodologies.