fastshap

Train a surrogate MLP that accepts masked features using Surrogate.train with a MarginalImputer, then train the explainer via FastSHAP.train with num_samples coalitions per batch. After training, call fastshap.shap_values(X_test) to get explanations in one forward pass.

KernelSHAP estimates Shapley values separately for each sample through repeated model evaluations, while FastSHAP trains a neural network to amortize this cost. Once trained, FastSHAP produces explanations in a single forward pass, making it suitable for real-time or batch explanation.

Yes, FastSHAP supports image data through ImageSurrogate and BaselineImageImputer, which mask superpixels of a fixed size. The explainer is typically a UNet that outputs per-superpixel Shapley values for each class.

MarginalImputer is recommended for tabular data because it replaces masked features with samples from the training distribution, staying faithful to the original model. BaselineImputer is faster but uses a fixed reference value and is less statistically principled.

Shapley values satisfy the efficiency axiom only when normalization is applied. Pass normalization='additive' to FastSHAP so predictions are adjusted such that their sum equals f(x) minus the null prediction.

Yes, the single-model approach trains the predictive model itself to handle missing features, removing the need for a surrogate. This requires retraining the original model, as shown in the CIFAR-10 single model notebook.