This is the code for our JMLR paper "Faith-Shap: The Faithful Shapley Interaction Index".

This code is tested on Python3.7 with the following packages:

• ktrain == 0.25.1
• tensorflow == 2.1.0
• transformers == 3.1.0
• xgboost == 1.6.2
• h5py == 2.10.0

• To replicate the computational efficiency experiment, run bash run_exp1.sh.
• solver.py contains our implementation of Faithful Shapley, shapley Taylor, and Shapley interaction indices.
• To compute Faith-Shap, we first sample each coalition $S \subseteq [d]$ with probability $\propto \frac{d-1}{{d \choose |S|}|S|(d-|S|)}$ in generate_perturbation.py, and solve the constrained regression.
• example.sh contains an example of applying faith-Shap to generate explanations on Bert trained on IMDB data.

If you find this code useful, please cite the following paper:

@article{tsai2023faith,
  title={Faith-shap: The faithful shapley interaction index},
  author={Tsai, Che-Ping and Yeh, Chih-Kuan and Ravikumar, Pradeep},
  journal={Journal of Machine Learning Research},
  volume={24},
  number={94},
  pages={1--42},
  year={2023}
}