sage-experiments

This repository contains code for the experiments in this paper. The code is frozen to use a specific implementation of SAGE, so if you want to use the current version please visit this repository.

Please send me an email if you find any problems here.

Replicating experiments

All of the datasets necessary to run our experiments are easily accessible except the breast cancer (BRCA) subtype classification dataset. Send me an email if you're interested in working with that data. To replicate the experiments described in the paper, follow these steps.

1. Running this code requires a number of Python packages (XGBoost, Catboost, PyTorch, etc.) in addition to the package for SAGE. You can install them all into your virtual environment with the following command:

pip install .

Code for SAGE can then be imported using the name sage.

2. In the experiments/ directory, run the train models.ipynb notebook to train and save models for all five datasets.

3. To generate global explanations using SAGE and the baseline methods, run the following notebooks in the experiments/ directory:

• sage explanations.ipynb
• feature ablation.ipynb
• permutation tests.ipynb
• mean importance.ipynb
• univariate predictors.ipynb

4. Run sampling algorithms for SHAP and SAGE while saving intermediate results by running these notebooks (also in the experiments/ directory):

• sage convergence.ipynb
• shap convergence.ipynb

5. Train models with random subsets of features by running the random subsets.ipynb in the experiments/ directory.

6. Generate figures like the ones in the text by running the following notebooks (also in the experiments/ directory):

• examples.ipynb
• more examples.ipynb
• model monitoring.ipynb
• cumulative correlation.ipynb
• feature selection.ipynb
• convergence.ipynb

References

Ian Covert, Scott Lundberg, Su-In Lee. "Understanding Global Feature Contributions With Additive Importance Measures." NeurIPS 2020.