Relevant Irrelevance: Generating Alterfactual Explanations for Image Classifiers

TL;DR

The paper tackles explainability for black-box image classifiers by introducing alterfactual explanations, which maximize changes to irrelevant input features while preserving the decision boundary and the classifier output. It presents a GAN-based framework capable of generating both alterfactual and counterfactual explanations for binary image classifiers, driven by four loss components including an auxiliary SVM-based relevance constraint. Evaluated on Fashion-MNIST, the approach achieves high alterfactual validity (≈96%) with low perceptual similarity (≈0.32) and strong counterfactual validity (≈88%) with high similarity (≈0.90), and a user study shows alterfactual explanations improve local model understanding and complement counterfactuals. The findings suggest that communicating irrelevance via alterfactual explanations can enhance users' mental models of AI decisions and inform broader XAI design, with potential applicability beyond Fashion-MNIST.”

Abstract

In this paper, we demonstrate the feasibility of alterfactual explanations for black box image classifiers. Traditional explanation mechanisms from the field of Counterfactual Thinking are a widely-used paradigm for Explainable Artificial Intelligence (XAI), as they follow a natural way of reasoning that humans are familiar with. However, most common approaches from this field are based on communicating information about features or characteristics that are especially important for an AI's decision. However, to fully understand a decision, not only knowledge about relevant features is needed, but the awareness of irrelevant information also highly contributes to the creation of a user's mental model of an AI system. To this end, a novel approach for explaining AI systems called alterfactual explanations was recently proposed on a conceptual level. It is based on showing an alternative reality where irrelevant features of an AI's input are altered. By doing so, the user directly sees which input data characteristics can change arbitrarily without influencing the AI's decision. In this paper, we show for the first time that it is possible to apply this idea to black box models based on neural networks. To this end, we present a GAN-based approach to generate these alterfactual explanations for binary image classifiers. Further, we present a user study that gives interesting insights on how alterfactual explanations can complement counterfactual explanations.

Figures (40)

• Figure 1: (A) Examples of a counterfactual and an alterfactual explanation. Input features to a fictional decision system to be explained are Income and Gender, whereas the former is relevant and the latter is irrelevant to the AI's decision on whether a credit is given or not. (B) Conceptual comparison of factual, counterfactual, semifactual, and alterfactual explanations.
• Figure 2: Architecture overview of the generator network.
• Figure 3: Architecture overview of the discriminator network.
• Figure 4: Example outputs of our system. It can be seen that alterfactual explanations change features that are irrelevant to the classifier, e.g., the color of the shoes or the width of the boot shaft, while counterfactual explanations change relevant features like the presence or absence of a boot shaft. From top to bottom the original images are a correctly classified ankle boot and sneaker, followed by two inputs incorrectly classified as ankle boot and sneaker.
• Figure 5: Left: Mean participant prediction accuracy of the AI's prediction by condition. The conditions containing alterfactual explanations outperformed all other conditions. Right: Mean understanding of the irrelevant and relevant features in our study. Error bars represent the 95% CI. *p$<$ .05, **p$<$ .001.