Gaussian Process Neural Additive Models

Authors: Wei Zhang, Brian Barr, John Paisley

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We experiment with several tabular data sets. A key feature of our model is its reduction in parameters and convex optimization, while still providing competitive performance as a neural additive model.
Researcher Affiliation Collaboration 1Columbia University, New York, NY, USA 2Capital One, New York, NY, USA {wz2363, jwp2128}@columbia.edu, brian.barr@capitalone.com
Pseudocode Yes Algorithm 1: GP-NAM for regression and classficiation
Open Source Code Yes The GP-NAM code for regression and classification can be found at https://github.com/Wei2624/GPNAM
Open Datasets Yes We experiment using several tabular data sets. This includes CA Housing1, FICO2, for which we follow the processing of Radenovic, Dubey, and Mahajan (2022). We also report performance on MIMIC-II3, MIMIC-II4, Credit5, Click6, Microsoft7, Year8 and Yahoo9, Churn10, Adult11, Bikeshare12 tabular data sets. For these, we follow the processing in Chang, Caruana, and Goldenberg (2021); Popov, Morozov, and Babenko (2019). We also consider our own processing of credit lending data sets LCD13 and GMSC13 More information about these data sets is shown in Table 2.
Dataset Splits Yes Table 2: Statistics from the tabular data sets used in binary classification (top) and regression (bottom) experiments. Dataset #Train #Val #Test
Hardware Specification No The paper mentions running experiments 'on GPU' and 'on our CPU' but does not specify exact models or types of hardware.
Software Dependencies No The paper states 'All models are implemented with Py Torch' but does not specify version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes Since each GP is one dimensional, we empirically found that S = 100 works well.