New Metric EPC Score Validated Against Human Explanations
A new paper on arXiv (ID: 2607.29614) introduces the EPC score, an improved version of the Explainability-Performance Coefficient designed to evaluate the quality of explanations from Explainable AI (XAI) systems. This score provides a way to assess explanation fidelity that isn't tied to specific models, balancing how features are chosen with maintaining model performance. The researchers tested the EPC score across different data types, such as tabular, text, and images, demonstrating its ability to uncover connections among network activations, data complexity, and explainer effectiveness. Impressively, the score showed a strong link to human-generated explanations, meaning higher EPC scores often matched human understanding. This research is vital for improving AI trustworthiness, especially in critical applications.
Key facts
- Paper ID: arXiv:2607.29614
- Introduces EPC score, an extension of the Explainability-Performance Coefficient
- Metric balances feature selection sparsity and preserved model performance
- Validated across tabular, text, and image modalities
- Uncovers dependencies among network activations, data dimensionality, and explainer performance
- Validated against human-based explanations
- Higher EPC scores align with human lexical similarity
- Addresses challenges in trustworthy XAI for high-risk domains
Entities
Institutions
- arXiv