# Evaluating Explanation Without Ground Truth in Interpretable Machine Learning

**Authors:** Fan Yang, Mengnan Du, Xia Hu  
**arXiv:** [1907.06831](https://arxiv.org/abs/1907.06831)  
**Date:** 2019-07-16  
**Summary page:** https://www.summarizepaper.com/en/paper/1907.06831v2/

## Key points

- Interpretable Machine Learning (IML) is important in real-world applications like autonomous cars and medical diagnosis
- Evaluating the quality of explanations in IML is challenging due to diverse scenarios and lack of ground truth data
- The article defines three aspects of explanation: generalizability, fidelity, and persuasibility
- It reviews existing methodologies for evaluating explanations across different tasks
- A unified evaluation framework is proposed, considering the needs of developers and end-users
- Open problems in evaluating explanations are discussed, along with limitations of current techniques
- Addressing these challenges helps researchers understand the benefits of explanations for human users.

## Summary

Interpretable Machine Learning (IML) has gained significant importance in various real-world applications, including autonomous cars and medical diagnosis. The need for explanations in these applications is crucial to help users understand how machine learning systems work and build trust in them. However, evaluating the quality of generated explanations in IML is challenging due to the diverse scenarios and subjective nature of explanations. Ground truth data for benchmark evaluation is rarely available. This article addresses this issue by rigorously defining the problem of evaluating explanations in IML and reviewing existing efforts from state-of-the-art research. The authors summarize three general aspects of explanation: generalizability, fidelity, and persuasibility. They provide formal definitions for each aspect and review representative methodologies for evaluating them across different tasks. Additionally, they propose a unified evaluation framework that considers the hierarchical needs of developers and end-users, making it adaptable to various practical scenarios. The article also discusses open problems in evaluating explanations and highlights limitations of current evaluation techniques that need further exploration. By addressing these challenges, researchers can better assess system boundaries and understand the true benefits that explanations offer to human users. In summary, this article provides a comprehensive overview of evaluating explanations without ground truth in Interpretable Machine Learning. It defines key aspects of explanation quality, reviews existing methodologies, proposes a unified evaluation framework, and identifies areas for future research.