I referred to this beautiful document to research about joint feature contibutions. But this works only for RandomForest algorithms because of treeinterpreter (does not work with xgboost). Is there a similar way out for XGBoost as well?

Basically what I want to achieve is to find out the joint contributions of all the combination of features towards the prediction. For example if I have a, b and c as my features, I want to know what is the effect of ab, bc and ca towards the prediction result. It is very similar to shap and lime, but for combinations of features.


1 Answer 1


I did some research and learnt about xgbfir package. It gives the joint contributions into an excel file. You can set the level of interaction with this. I wrote some code around it to generate a plot that solves the purpose.

If the package is not installed

pip install xgbfir

After the installation:

import xgbfir
from matplotlib import pyplot as plt

xgbfir.saveXgbFI(model, feature_names=X.columns, OutputXlsxFile='FI.xlsx')

joint_contrib = pd.read_excel('FI.xlsx')

xls = pd.ExcelFile('FI.xlsx')
df1 = pd.read_excel(xls, 'Interaction Depth 0')
df2 = pd.read_excel(xls, 'Interaction Depth 1')
df3 = pd.read_excel(xls, 'Interaction Depth 2')

frames = [df1, df2, df3]
joint_contrib = pd.concat(frames)

joint_contrib=joint_contrib.sort_values(by='Gain', ascending=True)

height = joint_contrib['Gain']
bars = joint_contrib['Interaction']
y_pos = np.arange(len(bars))

plt.barh(y_pos, height)
plt.yticks(y_pos, bars)

This will give the top 20 feature interactions in terms of gain.

Thanks to Philip Cho who introduced me to xgbfir.

Follow the link for more information regarding xgbfir


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.