Carefully choose the baseline: Lessons learned from applying XAI attribution methods for regression tasks in geoscience

Baseline choice fundamentally shapes XAI results: The paper demonstrates that different baselines (reference points used by explainability methods) can produce substantially different attribution results for the same neural network prediction, meaning there is no single "correct" explanation—the choice of baseline determines what insights you get.

Users often make baseline choices unknowingly: Many XAI methods set a default baseline automatically without making users aware of this choice, which can lead to misinterpretation of results. Practitioners need to understand what baseline their explainability tool is using and why.

Baseline selection should match your research question: Different baselines answer different scientific questions; the paper shows that in climate modeling, choosing different reference points reveals different aspects of how the neural network makes predictions, so baseline selection should be deliberate and aligned with what you're trying to understand.

Current XAI literature overlooks this critical issue: The authors note that the impact of baseline choice "has been overlooked so far in the literature," suggesting that many published XAI applications in geoscience (and potentially other fields) may not have adequately considered how their baseline choice affected their conclusions.

Responsible AI requires transparency about methodological choices: For AI systems to be trustworthy and interpretable, researchers must explicitly document and justify their baseline selection rather than relying on default settings, ensuring that stakeholders understand the assumptions underlying the explanations provided.