Skip to main content

KernelShap

LocalExplainer
DashAI.back.explainability.explainers.KernelShap

Model-agnostic local explainer that estimates SHAP values via a weighted linear model.

Kernel SHAP (SHapley Additive exPlanations) unifies LIME and classic Shapley values from cooperative game theory. For each instance to explain, it fits a weighted linear model over a sampled coalition of feature subsets, where the sample weights are derived from the Shapley kernel. The resulting coefficients are the SHAP values — each one represents the marginal contribution of a feature to the model's prediction relative to a background (reference) distribution.

Because it treats the model as a black box (querying only predict_proba), Kernel SHAP works with any classifier. The trade-off is higher computational cost compared to model-specific SHAP implementations (Tree SHAP, Deep SHAP).

References​

Parameters​

link : string, default=identity
Link function to connect feature importance values to the model's outputs. Options are 'identity' (identity function) or 'logit' (log-odds).
fit_parameter_sample_background_data : boolean, default=True
Parameter to fit the explainer. 'true' if background data must be sampled; otherwise the entire training set is used. Smaller datasets speed up the algorithm runtime.
fit_parameter_background_fraction : number, default=0.2
If 'Sample background data' is selected, this corresponds to the fraction of background samples to draw from the training set.
fit_parameter_sampling_method : string, default=shuffle
If 'true', choose to sample random instances with 'shuffle' or summarize the dataset with 'kmeans'. If there are categorical features, 'shuffle' is used by default.

Methods​

explain_instance(self, instances)

Defined on KernelShap

Method for explaining the model prediciton of an instance using the Kernel Shap method.

Parameters

instances: DatasetDict
Instances to be explained.

Returns

dict
dictionary with the shap values for each instance.

fit(self, background_dataset, sample_background_data='false', background_fraction=None, sampling_method=None)

Defined on KernelShap

Method to train the KernelShap explainer.

Parameters

background_data: Tuple[DatasetDict, DatasetDict]
Tuple with (input_samples, targets). Input samples are used to estimate feature attributions and establish a baseline for the calculation of SHAP values.
sample_background_data: bool
True if the background data must be sampled. Smaller data sets speed up the algorithm run time. False by default.
background_fraction: float
Proportion of background data from the training samples used to estimate SHAP values if sample_background_data=True.
sampling_method: str
Sampling method used to select the background samples if sample_background_data=True. Options are 'shuffle' to select random samples or 'kmeans' to summarise the data set. 'kmeans' option can only be used if there are no categorical features.

Returns

KernelShap object

plot(self, explanation: list[dict])

Defined on KernelShap

Method to create the explanation plot using plotly.

Parameters

explanation: dict
dictionary with the explanation generated by the explainer.
Returns:
List[dict]
list of JSONs containing the information of the explanation plot to be rendered.

get_schema(cls) -> dict

Defined on ConfigObject

Generates the component related Json Schema.

Returns

dict
Dictionary representing the Json Schema of the component.

validate_and_transform(self, raw_data: dict) -> dict

Defined on ConfigObject

It takes the data given by the user to initialize the model and returns it with all the objects that the model needs to work.

Parameters

raw_data : dict
A dictionary with the data provided by the user to initialize the model.

Returns

dict
A validated dictionary with the necessary objects.

Compatible with​