feature-tweaking

Implementation of the Feature Tweaking algorithm for XGBoost models

Interpretable Predictions of Tree-based Ensembles via Actionable Feature Tweaking, Tolomei et. al. https://arxiv.org/abs/1706.06691

Setup

The package may be installed by

pip installl git+https://github.com/nicholaskarlsen/feature-tweaking

Usage

# the feature-tweaking class may be instantiated in the following way
ft = FeatureTweaking(model, continuous_features, categorical_features, continuous_metric="l1", categorical_metric="kronecker delta")

# which may then be used to generate counterfactual examples
counterfactuals = ft.generate_counterfactuals(X.iloc[i], epsilon=0.1)
# the factual class is inferred from the passed model and a pandas dataframe contiaining the set of epsilon-satisfactory examples
# is returned along with their distance from the factual example measured using the selected metric

Parameter	Description
`model (xgboost.Booster	xgboost.sklearn.XGBModel)`
`continuous_features (list<str>)`	list containing column names of the continuous features to tweak
`categorical_features(list<str>)`	list containing column names of the categorical features to tweak
`continuous_metric (str)`	metric to use for all of the continuous features. see FeatureTweaking._metrics for a list of currently implemented metrics
`categorical_metric (str)`	metric to use for all of the categorical features. see FeatureTweaking._categorical_metrics for a list of currently implemented metrics