explainability.SeldonExplainer

class SeldonExplainer(estimator, background, *, index='k-SII', max_order=2, class_index=None, imputer='marginal', max_background_samples=100, feature_names=None, random_state=None, verbose=False)

Explain predictions of a fitted Seldon estimator with Shapley values.

Wraps shapiq.TabularExplainer to provide both local (per-instance) and global (dataset-level) explanations. For classifiers the explanations are computed on a single class probability selected by class_index; for regressors they are computed on the predicted value.

Parameters:

  • estimator (SeldonClassifier, SeldonRegressor or Seldon) – A fitted Seldon estimator. The Seldon wrapper is accepted and unwrapped to its inner estimator.

  • background (array-like of shape (n_samples, n_features)) – Reference data used to impute the values of features that are “absent” from a coalition. Typically (a subset of) the training data. To bound cost and latency, it is subsampled to max_background_samples rows.

  • index (str, default="k-SII") –

    Interaction index to estimate. Common choices:

    • "SV": plain Shapley values (no interactions, max_order forced to 1).

    • "k-SII": k-order Shapley Interaction Index (default).

    • "FSII": Faithful Shapley Interaction Index.

    Any index supported by shapiq.TabularExplainer is accepted.

  • max_order (int, default=2) – Maximum interaction order to estimate. 1 yields feature attributions only; 2 additionally captures pairwise interactions. Ignored (forced to 1) when index="SV".

  • class_index (int, optional) – For classification, the index into classes_ whose probability is explained. Defaults to the positive class (1) for binary targets; for multiclass targets it defaults to 1 with a warning — pass an explicit value to choose the class of interest. Ignored for regression.

  • imputer (str, default="marginal") – Imputation strategy for absent features, forwarded to shapiq (e.g. "marginal", "conditional", "baseline").

  • max_background_samples (int, default=100) – Maximum number of background rows to keep. Larger background sets give more faithful imputation but increase the size of every remote request.

  • feature_names (sequence of str, optional) – Names used when plotting. Defaults to the estimator’s column_names if set, otherwise None.

  • random_state (int, optional) – Seed controlling background subsampling and shapiq’s sampling.

  • verbose (bool, default=False) – Forwarded to shapiq to show a progress bar during estimation.

Variables:

  • task_type (str) – "classification" or "regression", copied from the estimator.

  • class_index (int or None) – Resolved class index being explained (classification only).

  • feature_names (list of str or None) – Feature names used for plotting.

  • background (ndarray of shape (n_kept, n_features)) – The (subsampled) background data passed to shapiq.

See also

neuralk.explainability.GlobalExplanation

Result of explain_global.

Examples

>>> explainer = SeldonExplainer(clf, background=X_train, index="k-SII")
>>> iv = explainer.explain(X_test[0])
>>> iv.plot_force(feature_names=explainer.feature_names)