Diffusion Counterfactuals for Image Regressors

Counterfactual explanations can now be applied to regression models, not just classification tasks. This is significant because many real-world AI systems (predicting ages, prices, or risk scores) use regression, yet explaining their decisions has been underexplored until this work.

The methods revealed that regression models require larger, more noticeable changes to features to produce meaningful shifts in predictions compared to classification models. This finding suggests regression explanations may be inherently harder to make sparse and simple, which has implications for how transparent these systems can realistically be to users.

The research demonstrates that diffusion-based generative models can create realistic, interpretable counterfactuals that show what would need to change in an image for a model to make a different prediction—offering a human-friendly way to understand AI decision-making in visual domains.

These counterfactual explanations can expose spurious correlations that models have learned, helping practitioners identify when AI systems are relying on irrelevant or problematic patterns rather than legitimate features—a critical step in responsible AI auditing.

There's a trade-off between explanation sparsity and quality: pixel-space explanations are simpler but lower quality, while latent-space explanations are more realistic but involve larger changes, suggesting practitioners must choose based on whether they prioritize simplicity or fidelity in explanations.