Generating and Testing Synthetic Datasets for Recommender Systems to Improve Fairness in Collaborative Filtering Research

Fairness is an important field in the Recommender Systems area since collaborative filtering datasets tend to be demographically biased. This paper proposes a parameterized model and provides an open repository of code to generate synthetic datasets containing demographic information. The model parameters can be set to hold different numbers of minority and nonminority users, distributions, and their overlapping. A neural network model has also been used to test the accuracy obtained in different scenarios, by setting the number of minority users and the overlapping between minority and nonminority distributions of votes. The results show how minority users receive unfair recommendations, particularly when their number decreases and when the distributions of minority versus nonminority users partially overlap. This research can be easily extended by designing more sophisticated models and modifying the provided framework; in this sense, a specific future work has been proposed.