How to pick the smartest questions when estimating feature importance is expensive?

Computing Shapley values—a gold standard for explaining AI decisions—requires evaluating thousands of feature combinations, which becomes prohibitive when each evaluation is expensive (retraining models, testing hyperparameters). ShaplEIG uses a Gaussian process to learn patterns from initial evaluations, then intelligently picks which combinations to test next by maximizing information gain. A mathematical trick using symmetric polynomials makes this feasible without exponential blowup. Experiments show it cuts sample costs substantially when budget is tight.