Explainability

Local and global explanations for fitted Seldon estimators, powered by the shapiq package.

Note

Explanations require the optional shapiq dependency (Python 3.12+):

pip install "neuralk[explain]"

Warning

Seldon predicts through remote API calls that consume credits. A single SeldonExplainer.explain call scores about budget x len(background_) perturbed rows (25,600 with the defaults) in a couple of requests, each re-sending the training set; SeldonExplainer.explain_global repeats that for every row of X. Keep background small and budget modest while iterating.

Build a SeldonExplainer from a fitted estimator: SeldonExplainer(clf, background=X_train).

SeldonExplainer(estimator, background, *[, ...])	Explain predictions of a fitted Seldon estimator with Shapley values.
GlobalExplanation(interaction_values, *[, ...])	Aggregated, dataset-level explanation produced by SeldonExplainer.explain_global.