Choosing the Baseline in Feature Attribution

Every feature-attribution method — SHAP, integrated gradients, LIME — answers a question of the form "how much did each feature contribute to this prediction, compared to some reference?" That reference, the baseline, represents "no information" or "a typical, uninformative input," and every reported attribution number is only meaningful relative to it. Change the baseline, and the attributions change too, sometimes dramatically, even though the model and the prediction being explained haven't moved at all.

This isn't a bug to be fixed; it's structurally unavoidable, because "how much did this feature contribute" only has a well-defined answer once you say "contribute, compared to what." Common choices include the dataset's average feature values (a "typical" input), an all-zeros vector (useful when zero genuinely means absence, less useful when it doesn't — a stock's zero price isn't a meaningful reference point), or a background sample drawn from the actual data distribution, which many implementations of SHAP use by default, effectively averaging attributions over many different plausible baselines at once. None of these is objectively "correct" — they encode different questions. "Average feature values" answers "how does this input differ from a typical one?" A background sample answers something closer to "how does this input differ from other inputs the model has actually seen?"

Worked example

A model predicts default probability from two features: debt-to-income ratio and account age. For one applicant with debt-to-income of 0.6 and account age of 2 years, the model predicts 30% default probability. Using the dataset average as the baseline (debt-to-income 0.35, account age 6 years, baseline prediction 12%), SHAP attributes the 18-point gap mostly to debt-to-income (say, +14 points) with a smaller contribution from account age (+4 points), because this applicant's debt-to-income is unusually far from the population's typical value. Using an all-zeros baseline instead (debt-to-income 0, account age 0, baseline prediction 5%), the 25-point gap now splits differently — account age contributes much more (+11 points), because going from zero years of history to two years is itself a meaningful change relative to a baseline of no account history at all, a comparison the average-based baseline never made since the typical applicant already has an account. Both attributions are internally consistent and mathematically valid; they simply answer different questions about what "no information" means.

baseline: all-zeros DTI: +14 age: +11

What this means in practice

Any reported feature attribution should come with a statement of what baseline was used, the same way a treatment-effect estimate should come with a statement of what population it applies to — "debt-to-income contributed +14 points" is incomplete without "...relative to the population average applicant." Comparing attributions produced with different baselines across two studies, two vendors, or two versions of a model can look like a real disagreement about what matters when it's actually just a difference in reference point.