Dynamic paired comparison modelling with Gaussian Processes

This repository contains code to fit dynamic paired comparison models using Gaussian Process priors, as described in this preprint:

Gaussian Process Priors for Dynamic Paired Comparison Modelling

Requirements

The minimum requirements to use the library are listed in requirements.txt. You can install them by running:

pip install -r requirements.txt

Note that scikit-sparse requires that the SuiteSparse library is installed. This dependency is most easily handled with anaconda, where you can do:

conda install suitesparse

If that doesn't work for you, there are other instructions here: scikit-sparse documentation

If you want to run the demo notebooks, you will also have to install the requirements in demo_requirements.txt:

pip install -r demo_requirements.txt

Once the requirements are installed, you can run:

python setup.py install

To install the library for you.

How to use

The easiest way to get started is to view the demos in the jupyter folder.

• Time only demo.ipynb fits a Matern 3/2 kernel to tennis data, shows the inferred latent functions, and has a prediction example.
• Surface demo.ipynb fits a Matern 3/2 kernel on time multiplied with an ARD RBF kernel on surface, shows the latent functions, and has a prediction example, too.
• Bayesian optimisation demo.ipynb shows an example of how to run Bayesian Optimisation to maximise the log marginal likelihood on the Matern 3/2 kernel.

Citing

If you would like to use this code in your academic work, please cite the paper below:

• Martin Ingram: "Gaussian Process Priors for Dynamic Paired Comparison Modelling", 2019; URL