This is the code for the paper Hodge-Compositional (HC) edge Gaussian processes (GPs) in AISTATS 2024. It contains the three experiments, together with the corresponding datasets, as well as the kernel classes.
create an environment and install the needed packages
conda env create -f hc-gp-environment.ymlWe provide Jupyter Notebook files for the three experiments.
In GPRegression_Forex, we provide the tutorial of using edge Gaussian processes to interpolate a foreign currency exchange (Forex) market. The dataset is forex_2018.pkl. The kernels are written using GPyTorch in edge_kernel_forex. One can select the following kernels implmented:
- HC Maten or diffusion kernel
- non-HC Matern or diffusion kernel
- Euclidean Matern or rbf kernel
- Line-graph Matern or diffusion kernel
Likewise, in GPRegression_OceanFlow and GPRegression_WSN, we provide the tutorials for interpolating the ocean edge flows and a water supply network.
For example, to run the Forex experiment with the HC edge Matern kernel with 20% training ratio, one can bash the following
run -i GPRegression_Forex.ipynb --seed 5 --kernel_name matern --train_ratio 0.2@InProceedings{pmlr-v238-yang24e,
title = {Hodge-Compositional Edge {G}aussian Processes},
author = {Yang, Maosheng and Borovitskiy, Viacheslav and Isufi, Elvin},
booktitle = {Proceedings of The 27th International Conference on Artificial Intelligence and Statistics},
pages = {3754--3762},
year = {2024},
editor = {Dasgupta, Sanjoy and Mandt, Stephan and Li, Yingzhen},
volume = {238},
series = {Proceedings of Machine Learning Research},
month = {02--04 May},
publisher = {PMLR},
pdf = {https://proceedings.mlr.press/v238/yang24e/yang24e.pdf},
url = {https://proceedings.mlr.press/v238/yang24e.html}
}
The project is listed under the MIT license.