Symbolic Pursuit

Github for the NIPS 2020 paper "Learning outside the black-box: at the pursuit of interpretable models"

Installation

The library can be installed from PyPI using

$ pip install symbolic_pursuit

or from source, using

$ pip install .

Example Usage

To build a symbolic regressor for a given dataset and a given model (or a given model type), the following command can be used :

 python3 build_interpreter.py [-h] [--dataset DATASET] [--test_ratio TEST_RATIO]
                            [--model MODEL] [--model_type MODEL_TYPE]
                            [--verbosity VERBOSITY] [--loss_tol LOSS_TOL]
                            [--ratio_tol RATIO_TOL] [--maxiter MAXITER]
                            [--eps EPS] [--random_seed RANDOM_SEED]

For example, if one would like to train a MLP one the wine-quality-red dataset and then fit a symbolic regressor with random seed 27, one can use the command

python3 build_interpreter --dataset wine-quality-red --model_type MLP --random_seed 27

For more details on how to use the module in general, see the 3 enclosed notebooks.

1. Building a Symbolic Regressor 2. Symbolic Pursuit vs LIME 3. Synthetic experiments with Symbolic Pursuit

🔨 Tests

Install the testing dependencies using

pip install .[testing]

The tests can be executed using

pytest -vsx

References

In our experiments, we used implementations of LIME, SHAP and pysymbolic

Citing

If you use this code, please cite the associated paper:

@article{https://doi.org/10.48550/arxiv.2011.08596,
  doi = {10.48550/ARXIV.2011.08596},
  url = {https://arxiv.org/abs/2011.08596},
  author = {Crabbé, Jonathan and Zhang, Yao and Zame, William and van der Schaar, Mihaela},
  keywords = {Machine Learning (cs.LG), Artificial Intelligence (cs.AI), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {Learning outside the Black-Box: The pursuit of interpretable models},
  publisher = {NeurIPS 2020},
  year = {2020},
}