prefGP

prefGP is a Gaussian process based library for learning from preference and choice data. It is implemented using Jax and PyTorch. prefGP implements 9 models to learn from preference and choice data:

• Model 1: Consistent Preferences.
• Model 2: Just Noticeable Difference
• Model 3: Probit for Erroneous Preferences
• Model 4: Preferences with Gaussian noise error
• Model 5: Probit for Erroneous preferences as a classification problem
• Model 6: Thurstonian model for label preferences
• Model 7: Plackett-Luce model for label ordering data
• Model 8: Paired comparison for label preferences
• Model 9: Rational and Pseudo-rational models for choice data

Requirements:

• Python == 3.11.4

Download the repository and then install

pip install -r requirements.txt

The notebooks folder includes several ipython notebooks that demonstrate the use of prefGP. For more details about the models used in the examples, please see the below paper.

@article{prefGP2024,
  title = {A tutorial on learning from preferences and choices with Gaussian Processes},
  author = {Benavoli, Alessio and Azzimonti, Dario},
  journal = {arXiv preprint},
  year = {2024},
  eprint = {2403.11782},
  url = {https://arxiv.org/abs/2403.11782}
}

The library was developed by

• Dario Azzimonti (IDSIA)
• Alessio Benavoli (Trinity College Dublin)