PatternGAM

Code for the paper "Correcting misinterpretations of additive models", accepted to NeurIPS 2025: https://openreview.net/forum?id=2ClM0g9OFT.

The ./nam/ subfolder is adapted from lemeln/nam.

Conda Setup

Set up the environment using the provided environment.yml:

# Create the environment
conda env create -f environment.yml

# Activate the environment
conda activate <env_name>

To generate XAI-TRIS data and save it to ./artifacts/tetris/data/neurips, run:

python -m data.generate_data data_config 8x8 neurips

This uses scenarios from ./data/data_config.json.
generate_linesearch.py creates datasets for SNR-tuning line search, which is then run in linesearch.py.
visualise_linesearch.ipynb generates training result plots.
xai_tris_train_models.py trains the final model parameterizations. Ensure the data folder path at the top of each script matches your setup (e.g., ./artifacts/tetris/data/...).
xai_tris_shape_fns.ipynb shows shape functions and the interaction plots for XAI-TRIS
xai_tris_explanations.py generates global explanations for the notebooks:
- xai_tris_qualitative.ipynb
- xai_tris_quantitative_metrics.ipynb
These notebooks generate the plots shown in the paper.
fni_emd.ipynb demonstrates explanation styles in the appendix and their metric scores.

recid.data is used for experiments.
compas_nam.ipynb trains the NAM and generates shape functions and PatternGAM results.
compas_correlation.py creates correlation plots for the appendix.

Original data and related articles: propublica/compas-analysis.

mimic_preprocessing.py processes raw MIMIC-IV data for the 24-hour mortality task (tested with v2.0).
mimic_nam.py runs the experiments shown in the main text. Reduce epochs or learners for quicker results.

Data source: MIMIC-IV v2.0 (requires training, data use agreement, and access request).