Use of conditional generative adversarial networks to create demographic collaborative filtering datasets

This paper proposes a method to create synthetic collaborative filtering datasets that can be used to test both current and new fair recommender systems models. The proposed "Conditional Generative Adversarial Network for Recommender Systems (CGANRS)" method generalizes the existing generative adversarial network for recommender systems one, and it makes use of a conditional generative adversarial network to artificially generate synthetic profiles from a source dataset such as MovieLens. The created datasets can be parameterized to have different sizes and to include different number of users and items. Additionally, the provided parameters include the proportion of multi-categorical demographic information such as the number of male vs. female users, or the proportions of very young, young, adult, and senior users. To test the proposed method, three sets of synthetic databases have been created, containing different a) numbers of users, b) numbers of items, and c) proportions of male users versus female users. Results show an adequate behavior of the generated datasets, testing their a) profiles separability, b) main statistical distributions, and c) recommendation accuracies. Synthetic data sets created using the proposed conditional generative adversarial network for recommender systems method are particularly useful to improve research in the fairness field of the recommender systems area. To extend its use and to facilitate reproducibility, the source code is provided to generate as many demographic datasets as desired, as well as the artificially generated datasets in this research. Some promising future works are proposed, including a) the variation of the stochastic Gaussian distribution used to create the random noise vectors that feed the adversarial network generator model, and b) testing the fairness of the most relevant collaborative filtering models on different synthetic scenarios.