AlibiModelExplainer¶

This class provides methods to explain machine learning models using Alibi. It is a subclass of BaseModelExplainer.

Configuration¶

Required Configuration¶

The Alibi model explainer requires the following configuration:

local_dir: Location of a local directory to output files generated by this component.
explainer_cl: The alibi.explainer class to user to generate explanations.

Optional Configuration¶

The Alibi model explainer uses no optional configuration:

Default Configuration¶

The Alibi model explainer uses the following optional configuration:

skip_explainer_plots: Whether or not to generate explainer plots. Defaults to False.

Methods¶

get_explanations¶

Generates feature importance scores for a given model and input data.

def get_explanations(data, model, model_version, label, classification_type=None, to_list=False, skip_explainer_plots=True, *args, **kwargs)

Arguments:

data (pandas.DataFrame): The input data to explain the model.
model (object): The machine learning model to explain.
model_version (string): The model version obtained from the metadata_tracker.
label (string): The model target.
classification_type (string): The type of classification. This parameter is optional and only used if the problem type is classification.
to_list (boolean): If True, the returned explanations will be in a list format.
skip_explainer_plots (boolean): If True, skip the explainer plots.

Returns:

A dictionary containing explainer values.

get_feature_importance¶

Generates feature importance scores for train and test sets.

def get_feature_importance(data_dict, model, model_version, *args, **kwargs)

Argumentss:

data_dict (dict): A dictionary of the input data to explain.
model (object): The machine learning model to explain.
model_version (string): The model version obtained from the metadata_tracker.

Returns:

Tuple of two Alibi data objects: - the feature importance for the training set. - the feature importance for the test set.

_get_shap_plots¶

Generates various SHAP plots for the given SHAP values and saves them as artifacts.

def _get_shap_plots(shap_values, expected_value, data, model, label, model_version, classification_type=None)

Arguments:

shap_values (numpy.ndarray): The SHAP values to generate the plots.
expected_value (float): The expected value of the model.
data (pandas.DataFrame): The input data to the model.
model (object): The machine learning model to explain.
label (string): A label for the data.
model_version (string): The model version.
classification_type (string): The type of classification.

_save_pyplot¶

Saves the generated plot as an artifact.

def _save_pyplot(name, label, model_version)

Arguments:

name (string): A name for the plot.
label (string): The label of the data.
model_version (string): The model version obtained from the metadata_tracker.

Returns: The saved artifact.

_save_file¶

Saves the generated plot to a file as an artifact.

def _save_file(name, label, content, model_version, extension="html")

Arguments:

name (string): A name for the file.
label (string): The label of the data.
content (string): The content of the file.
model_version (string): The version of the model to explain.
extension (string): The extension of the file.

Returns:

The saved artifact.

_compare_train_test_feat_importance¶

Compares the feature importance scores between the train and test sets.

def _compare_train_test_feat_importance(explanations_train, explanations_test, classification_type, threshold=0.25)

Arguments:

explanations_train (pandas.DataFrame): The SHAP-based feature importance for the training set.
explanations_test (pandas.DataFrame): The SHAP-based feature importance for the test set.
classification_type (string): The type of classification.
threshold (float): A threshold value for checking the differences between train and test sets.

Returns:

Tuple containing the expected differences and the feature importance differences.

Usage¶

from lolpop.explainer import AlibiModelExplainer

... ## generate data, model, and model_verison 

config = {
    #insert component configuration
}

explainer = AlibiModelExplainer(conf=config)


# Generate SHAP-based feature importance scores for train and test sets
explanations_train, explanations_test = explainer.get_feature_importance(data_dict, model, model_version)