This repository contains the code used to obtain results for our arXiv paper "ParamBoost: Gradient Boosted Piecewise Cubic Polynomials".
We used uv as Python package manager. To obtain the benchmark results, simply run:
uv run benchmark_exp.py
Note that due to conflicts with python packages, EBM needs to be run separately in its own environment. This can be done, for example, with mamba:
mamba create -n interpret python==3.13
mamba activate interpret
pip install -r requirements.txt
Then run:
python benchmark_exp_ebm.py
Where the lines from benchmark_exp.py that would result in ImportError have been commented out for convenience.
To obtain the case study results, simply run:
uv run case_study/lpmc.py
Due to some numerical errors, the results may vary slightly from one run to another, but should be consistent overall.
Note that the datasets need to be split in train--val--test sets before running the benchmark/case study scripts. To obtain new splits, simply run split_dataset.py.
Finally, note that some datasets had to be omitted because they were too large to be pushed on GitHub. They can be downloaded from the following source:
- MSD: https://archive.ics.uci.edu/dataset/203/yearpredictionmsd
- Diabetes: https://github.com/vduong143/CAT-KDD-2024/tree/main
- Fraud: https://www.kaggle.com/datasets/mlg-ulb/creditcardfraud
Then they need to be put in the data folder. The loading paths in split_dataset.py might need to be updated.
