Legacy Activation Migration

Case Explainer 0.2 supports the legacy activation constructor parameters while emitting DeprecationWarning. Version 0.3.0 will remove activation_extractor, activation_layer, blend_alpha, and use_output_weights from case_explainer.CaseExplainer.

Replace the shorthand constructor:

explainer = CaseExplainer(
    X_train,
    y_train,
    model=mlp,
    activation_layer="last_hidden",
    blend_alpha=0.25,
)

with an explicit similarity strategy:

explainer = CaseExplainer(
    X_train,
    y_train,
    similarity=Blend(
        HiddenActivations(model=mlp, layer="last_hidden"),
        features=0.25,
    ),
)

For a custom extractor, migrate its behavior into an explicit similarity strategy before upgrading to 0.3.0:

explainer = CaseExplainer(
    X_train,
    y_train,
    similarity=Blend(
        CustomActivations(
            model=model,
            extractor=CallableActivationExtractor(custom_activation_function),
        ),
        features=0.25,
    ),
)

CustomActivations accepts any ActivationExtractor implementation and, wrapped in Blend, preserves the existing feature/activation blend geometry without relying on deprecated constructor parameters.

Renamed retrieval= parameter

The retrieval= parameter and the *Retrieval configuration classes introduced in 0.2 are deprecated in favor of similarity= and the intent-named strategy classes. The mapping is direct:

Deprecated	Preferred
retrieval=HiddenActivationRetrieval(...)	similarity=HiddenActivations(...)
retrieval=CustomActivationRetrieval(...)	similarity=CustomActivations(...)
retrieval=TreeLeafRetrieval(...)	similarity=TreeLeaf(...)
retrieval=ForestProximityRetrieval(...)	similarity=ForestProximity(...)
...Retrieval(..., blend_alpha=a)	similarity=Blend(strategy, features=a)

The use_output_weights field is replaced by output_weighting, where use_output_weights=False becomes output_weighting="none".