gpnam 0.0.20

GPNAM - an interpretable Gaussian Process NAM model.

GPNAM: Gaussian Process Neural Additive Models

The framework of GPNAM. $z_s, c_s$ and the sinusoidal function are predefined from the paper and do not require training. The only trainable parameter is W that maps to the output of each shape function.

This repository contains the source code for the paper Gaussian Process Neural Additive Models that appears at AAAI 2024.

Basically, the GPNAM constructs a Neural Additive Model (NAM) by a GP with Random Fourier Features as the shape function for each input feature, which leads to a convex optimization with a significant reduction in trainable parameters.

Sklearn interface

You can install the package by:

pip install gpnam

Then, you can run the model simply by:

from gpnam.sklearn import GPNAM

"""
input_dim: the dimensions of input data
problem: type of the model, 'classification' or 'regression'
"""
gpnam = GPNAM(input_dim, problem)
gpnam.fit(X, y)

y_pred = gpnam.predict(X_test)

Build on your own

Data sets preparation

You can download the data sets locally:

python ./gpnam/download_datasets.py LCD GMSC CAHousing

Setup

Experiments

Then, you can run the experiment:

python main.py --dataset LCD --optimizer Adam --n_epochs 200

Matlab

Note: Statistics and Machine Learning Toolbox is required.

Citation

If you find this repo useful, please consider citing our paper:

@inproceedings{,
  title={Gaussian Process Neural Additive Models},
  author={Wei Zhang and Brian Barr and John Paisley},
  booktitle={AAAI Conference on Artificial Intelligence},
  year={2024}
}