Feature importance explainers¶
ConstantExplainer¶
xaib.explainers.feature_importance.constant_explainer.ConstantExplainer
Returns the predefined constant vector for every input. This baseline is useful for checking stability metrics. For example it should maximize Model randomization check
LimeExplainer¶
xaib.explainers.feature_importance.lime_explainer.LimeExplainer
Another popular feature importance approach is LIME. The concept is that it tests what happens to the predictions of the model when it receives variations of data. LIME generates a new dataset that consists of perturbed samples and predictions of the black box model on them. On this new dataset LIME trains an interpretable model, which is weighted by the proximity of the sampled instances to the instance of interest. The learned model should be a good approximation of the machine learning model predictions locally and explain original model’s predictions properly.
Source: LIME
LinearRegressionExplainer¶
xaib.explainers.feature_importance.linear_regression_explainer.LinearRegressionExplainer
This explainer gives results by the multiplying the input by coefficients of a trained linear regression.
RandomExplainer¶
xaib.explainers.feature_importance.random_explainer.RandomExplainer
Good baseline to represent average results of an explainer. Output distribution can be shifted and scaled to test for normalization.
ShapExplainer¶
xaib.explainers.feature_importance.shap_explainer.ShapExplainer
The shap method is considered as a method that can be applied to explain individual predictions of any ML model. Shap is based on a game theory approach to additively accumulate the contribution of all features involved in a model. This method assigns each feature an “importance value” within a set of conditional expectations for a particular prediction. The results of this additive procedure are called “shap values”. These values can be spread across from a “base value” (which represents the average of the observations). As one can see, since the shap method accounts for all features and randomness of their order, this method can be quite computationally expensive for large models.
Source shap