🤖 machine learning

A Human-Centered Validation of the Explainability-Performance Coefficient

This paper introduces the EPC score, a model-agnostic metric that quantifies explanation quality by balancing feature sparsity and performance preservation, and validates its effectiveness across multiple data modalities by demonstrating its strong alignment with human-centered understanding.

Original authors: Christian Oliva, Luis F. Lago-Fernández

Published 2026-08-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Christian Oliva, Luis F. Lago-Fernández

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you've built a super-smart robot that can predict the future, diagnose diseases, or decide who gets a loan. It's incredibly good at its job, but it's also a "black box"—you can't see how it thinks. It just gives you an answer, and you have to trust it. This is the world of Deep Learning, a branch of artificial intelligence where computers learn by mimicking the human brain. But here's the catch: because these robots are so complex, we often don't know why they make certain decisions. This is a big problem, especially if the robot makes a mistake or if a law says we have a right to know how decisions are made.

To fix this, scientists created a field called Explainable AI (XAI). Think of XAI as a translator that tries to explain the robot's thinking to us. It highlights the "clues" the robot used to make its choice, like pointing to a specific word in a sentence or a specific part of a photo. But here's the tricky part: sometimes the translator lies. It might point to a clue that looks important but actually has nothing to do with the decision, or it might miss the real clue entirely. We need a way to test if these explanations are actually telling the truth or just making things up.

This paper is about building a better "lie detector" for these AI explanations. The authors, Christian Oliva and Luis F. Lago-Fernández, introduce a new tool called the EPC Score. Imagine you are trying to solve a mystery, and you have a list of suspects. A good explanation would tell you exactly which few suspects are guilty and which ones are innocent. The EPC Score tests this by playing a game: it asks, "If we remove the clues the AI says are important, does the robot stop working? And if we remove the clues it says are unimportant, does the robot keep working just fine?" If the explanation is honest, the robot should crash when you take away the important clues and stay calm when you take away the junk. The EPC Score measures how well an explanation passes this test, giving it a single number to show how trustworthy it is.

The researchers didn't just make up a theory; they put their new score to the test across three very different worlds: numbers (like bank loan data), pictures (like recognizing digits or complex images), and words (like figuring out if a movie review is happy or sad). They pitted different explanation methods against each other, from simple ones that just look at the math to complex ones that try to guess the robot's mind without seeing inside it.

What they found is a bit like a race where the heavyweights stumbled and the sprinters won. The complex, "model-agnostic" methods (which try to work on any robot without knowing its insides) turned out to be incredibly slow and often gave poor explanations, especially when the data got complicated. On the other hand, a method called Integrated Gradients (IG) consistently came out on top. It was fast, accurate, and gave the highest EPC scores across the board.

But the real magic happened when they checked these scores against human intuition. In the text world, they compared the AI's "important words" against a dictionary of human feelings. They found that the explanations with the highest EPC scores were the ones that actually picked words that humans would agree are positive or negative. In the image world, they checked if the AI was looking at the right part of a picture (like a dog's face) versus just the background. Again, the high-scoring explanations were the ones that focused on the actual object, not just the scenery.

One of the most interesting discoveries was how the way you "erase" information changes the results. When they tested image explanations, they found that simply turning pixels black (zero-masking) created weird, fake edges that tricked the AI. It made some explanations look good just because they avoided these fake edges. But when they used a "blur" instead of blacking out pixels, the results changed completely. Suddenly, the explanations that looked at every tiny pixel (like Integrated Gradients) were the clear winners, while the ones that just looked at big, blurry blobs fell behind. This suggests that to get a true reading of an AI's mind, you have to be careful not to introduce fake clues while you're testing it.

In the end, the paper suggests that the EPC Score is a reliable way to tell a good explanation from a bad one. It proves that the best explanations aren't just about highlighting a general area; they are about pinpointing the exact, necessary evidence the AI needs to make its decision. While the authors don't claim this solves every mystery in AI, their experiments strongly suggest that if you want an explanation you can trust, you should look for the one that scores high on the EPC test, and that method, Integrated Gradients, seems to be the current champion in this race.

Technical Summary: A Human-Centered Validation of the Explainability-Performance Coefficient

Problem Statement
The rapid adoption of Deep Learning (DL) in high-risk domains has intensified the need for trustworthy Explainable Artificial Intelligence (XAI). However, objectively evaluating the fidelity of explanations remains a critical challenge. Existing evaluation metrics often rely on proxy measures such as fidelity, stability, or robustness, which capture only partial aspects of explanation quality. Crucially, many current metrics fail to explicitly account for the trade-off between the sparsity of selected features and the preservation of the model's predictive performance. Furthermore, there is a lack of consensus on whether high-scoring explanations align with human-centered understanding, particularly across diverse data modalities (tabular, text, and image) and network architectures.

Methodology
The authors propose a model-agnostic metric, the EPC score, which serves as a scalar summary of the Explainability-Performance Coefficient (EPC). The methodology is grounded in the premise that a meaningful explanation must not only highlight relevant features but also preserve the model's predictive capability when restricted to those features.

Theoretical Framework: The EPC quantifies the "performance gap" between two perturbation regimes based on a relevance vector $R_x$ generated by an explainer:
- Deletion Top-k: Nullifying the $k\%$ of features deemed most relevant. A faithful explanation should cause a rapid drop in model confidence.
- Deletion Bottom-k: Nullifying the $k\%$ of features deemed least relevant. A robust explanation should maintain model confidence, as non-essential features are removed.
  The EPC curve plots this performance gap against the sparsity level $k$ . The EPC score is derived by calculating the area between the observed EPC curve and an optimal linear reference ceiling ( $f(k) = k/100$ ), normalized to produce a single scalar value where higher scores indicate better alignment with the ideal trade-off.
Experimental Setup: The framework was validated across three data modalities:
- Tabular Data: Bank loan default dataset (Kaggle).
- Image Data: MNIST (digit recognition) and ImageNet (via Imagenette subset).
- Text Data: IMDB sentiment analysis.
The study evaluated a diverse suite of local explainers, including model-agnostic methods (SHAP, LIME) and model-specific methods (Gradient $\times$ Input, LRP, Integrated Gradients, Grad-CAM). A critical component of the methodology involved testing different nullification strategies ( $\Phi$ ). For image data, the authors compared standard zero-masking against Gaussian blurring to mitigate out-of-distribution artifacts that could bias visual evaluations.
Human-Centered Validation: To assess alignment with human understanding, the EPC scores were correlated with independent human-annotated resources:
- Text: The AFINN sentiment lexicon, measuring the agreement between word relevance rankings and human-curated sentiment polarity.
- Image: ImageNet Object Localization annotations (Regions of Interest), measuring the spatial precision of explanations against human-identified object boundaries.

Key Results

Metric Performance: The EPC score successfully discriminates between high-quality and superficial explanations. Across all modalities and architectures (MLPs, CNNs, LSTMs), Integrated Gradients (IG) consistently achieved the highest EPC scores, outperforming simpler gradient-based methods and model-agnostic approaches.
Computational Feasibility: Model-agnostic methods (SHAP, LIME) exhibited prohibitive computational costs in high-dimensional spaces (e.g., ImageNet), leading to their exclusion from complex benchmarks. Gradient-based methods were significantly more efficient.
Impact of Nullification Strategy: In image classification, the choice of nullification operator significantly altered results. Zero-masking favored coarse-grained methods like Grad-CAM due to the introduction of artificial edges. Conversely, Gaussian blurring eliminated this structural bias, revealing that pixel-level explainers (specifically IG) possess superior fidelity, achieving significantly higher EPC scores than coarse regional methods.
Human Alignment:
- Text: Explainers with higher EPC scores showed stronger Pearson correlation with the AFINN sentiment lexicon. IG achieved the highest correlation ( $\rho = 0.65$ ), confirming that the metric aligns with human lexical sentiment judgments.
- Image: Higher EPC scores correlated with superior precision in identifying human-annotated Regions of Interest (ROIs) at fine-grained thresholds (Top@1). The study revealed that while object localization (ROI) and explanation fidelity are related, they are not equivalent; pixel-level methods identified the necessary evidence within an ROI more effectively than coarse methods, which often highlighted dispensable background context.

Significance and Claims
The paper claims that the EPC score provides a robust, standardized, and quantitative framework for comparing XAI methods across diverse architectures and data types. Its primary significance lies in its human-centered validation: the authors demonstrate that the EPC score is not merely an abstract mathematical construct but strongly correlates with independent human judgments (lexical sentiment and spatial annotations).

The work highlights that explanation fidelity is intrinsically tied to data modality, network architecture, and the operational conditions of the explainer (e.g., activation functions and nullification strategies). Specifically, the authors argue that:

Integrated Gradients is the most robust explainer across the evaluated benchmarks.
Evaluation Bias: Visual fidelity evaluations are highly sensitive to perturbation strategies; Gaussian blurring is necessary to fairly evaluate pixel-level methods.
Complementary Dimensions: Object localization (where the evidence is) and explanation fidelity (which parts of the evidence are necessary) are complementary, not equivalent, dimensions of explainability. The EPC score effectively captures the latter, extending validation beyond simple object localization to the necessity of specific features for model prediction.

The authors conclude that the EPC score serves as a reliable diagnostic tool for identifying faithful explanations, though they note that future work should explore generative inpainting techniques for nullification and extend the metric to Large Language Models (LLMs) and Vision-Language Transformers.

More like this