Ludovico Boratto1 and Mirko Marras2
1 EURECAT, Spain
2 EPFL, Switzerland
A Python introductory toolbox for playing with bias and unfairness issues in personalized rankings.
Install Python (>=3.6):
sudo apt-get update
sudo apt-get install python3.6
Clone this repository:
git clone https://github.com/biasinrecsys/wsdm2021.git
Install the requirements:
cd wsdm2021
pip install -r requirements.txt
Notebook 1: Setup and run of a recommender system [Colab link]
In this notebook, we become familiar with the Python recommendation toolbox, in the simplest possible way. First, we setup the working environment in GDrive. Then, we go through the experimental pipeline, by:
- loading the Movielens 1M dataset;
- performing a train-test splitting;
- creating a pointwise / pairwise / random / mostpop recommendation object;
- training the model (if applicable);
- computing the user-item relevance matrix;
- calculating some of the recommendation metrics (e.g., NDCG, Item Coverage, Diversity, Novelty).
The trained models, together with the partial computation we will save (e.g., user-item relevance matrix or metrics), will be the starting point of the investigation and the treatment covered by the other Jupyter notebooks.
Notebook 2: Popularity bias in personalized rankings [Colab link]
This notebook will outline a short study of item popularity in recommender systems. We assume that the number of ratings is a proxy of the popularity of the item. First, we will compare the characteristics of the items recommended by pairwise, pointwise, random and mostpop strategies. Then, we will show how to setup and perform a post-processing mitigation approach against popularity.
Notebook 3: Item provider unfairness in personalized rankings [Colab link]
This notebook will consider the directors of movies in Movielens 1M as the item providers and investigates how unfairness based on gender groups affects providers' group visibility and exposure with respect to their representation in the item catalog. Then, we introduce a pre-processing strategy that upsamples interations involving items of a minority group to improve fairness, while resulting in a small loss in utility.
This code is provided for educational purposes and aims to facilitate reproduction of our results, and further research in this direction. We have done our best to document, refactor, and test the code before the tutorial.
If you find any bugs or would like to contribute new models, training protocols, etc, please let us know.
Please feel free to file issues and pull requests on the repo and we will address them as we can.
If you find our code useful in your work, please reference the studies that led to them:
Notebook 1:
@inproceedings{boratto2019effect,
title={The effect of algorithmic bias on recommender systems for massive open online courses},
author={Boratto, Ludovico and Fenu, Gianni and Marras, Mirko},
booktitle={European Conference on Information Retrieval},
pages={457--472},
year={2019},
organization={Springer}
}
Notebook 2:
@article{boratto2021102387,
title={Connecting user and item perspectives in popularity debiasing for collaborative recommendation},
author={Ludovico Boratto and Gianni Fenu and Mirko Marras}
journal={Information Processing & Management},
volume={58},
number={1},
pages={102387},
year={2021},
issn={0306-4573},
doi={https://doi.org/10.1016/j.ipm.2020.102387},
organization={Elsevier}
}
Notebook 3:
@article{boratto2020interplay,
title={Interplay between Upsampling and Regularization for Provider Fairness in Recommender Systems},
author={Boratto, Ludovico and Fenu, Gianni and Marras, Mirko},
journal={arXiv preprint arXiv:2006.04279},
year={2020}
}
This code is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
This software is distributed in the hope that it will be useful, but without any warranty; without even the implied warranty of merchantability or fitness for a particular purpose. See the GNU General Public License for details.
You should have received a copy of the GNU General Public License along with this source code. If not, go the following link: http://www.gnu.org/licenses/.