Relevant Irrelevance: Generating Alterfactual Explanations for Image Classifiers

Silvan Mertes, Tobias Huber, Christina Karle, Katharina Weitz, Ruben Schlagowski, Cristina Conati, Elisabeth AndréIJCAI 2024

Introduction

Most Explainable AI (XAI) research focuses on identifying which features are most relevant to a model's decision. Techniques like SHAP or LIME highlight pixels that drive a classification, while counterfactual explanations identify the minimal changes required to flip a label. However, this paper argues that understanding a model's logic requires knowing what is irrelevant just as much as what is relevant. To address this, the authors introduce "alterfactual" explanations—alternative instances where irrelevant features are changed maximally while the AI's decision remains identical.

The Need for Alterfactuals

Traditional XAI often fails to explicitly show the "volume" of irrelevant space. For example, in a fair credit-scoring model, it is crucial to verify that changing a person's gender or skin color does not alter the outcome. While counterfactuals define the boundary of a class (the "what if" of a different outcome), alterfactuals define the breadth of a class (the "what if" of the same outcome under different conditions).

Taxonomy of explanations: Alterfactuals show maximal change within the same class, while counterfactuals show minimal change to flip the class.

As shown in the figure above, an alterfactual aims for low similarity to the original input while maintaining the same classification. This "relevant irrelevance" helps users build a more robust mental model of the classifier by demonstrating exactly which features the model ignores.

Methodology: A GAN-Based Approach

The authors propose a Generative Adversarial Network (GAN) architecture to produce these explanations. The generator is guided by a multi-component loss function:

  1. Adversarial Loss: Ensures generated images remain realistic.
  2. Classification Loss: Forces the model to maintain the original class for alterfactuals or flip it for counterfactuals.
  3. SSIM (Similarity) Component: Unlike counterfactuals, which seek high similarity (minimal change), alterfactuals optimize for low similarity to prove that the altered features are truly irrelevant.
  4. Feature Relevance Component: An auxiliary SVM trained on the classifier's penultimate layer ensures the explanation does not accidentally drift toward a decision boundary.

Empirical Evaluation

The framework was primarily evaluated on the Fashion-MNIST dataset, specifically a binary task distinguishing between ankle boots and sneakers. The GAN successfully generated valid alterfactuals—maintaining the correct class—96.20% of the time.

Visual examples of original, alterfactual, and counterfactual images for Fashion-MNIST classification.

In a user study with 131 participants, the authors tested whether these explanations improved human understanding. Participants were asked to predict the AI's decision on new images. Those provided with a combination of alterfactual and counterfactual explanations performed significantly better than those with counterfactuals alone. The improvement showed a large effect size ($d = 0.97$), suggesting that seeing both what matters and what doesn't is the most effective way to understand AI behavior.

Conclusion and Future Work

Alterfactuals provide a powerful tool for local model understanding. By changing irrelevant features maximally, they offer high-confidence proof of a model's invariance to certain attributes. Future work includes extending this architecture to multi-class problems and testing the approach in expert domains, such as medical imaging, where users lack prior intuition about feature relevance.

Summary updated

BibTeX
@inproceedings{mertes2024relevant,
  author = {Silvan Mertes and Tobias Huber and Christina Karle and Katharina Weitz and Ruben Schlagowski and Cristina Conati and Elisabeth André},
  title = {{Relevant Irrelevance: Generating Alterfactual Explanations for Image Classifiers}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure