Cluster-based quotas for fairness improvements in music recommendation systems

Abstract

Increasingly, music recommendations are influencing user listening behavior, which naturally impacts the music industry, as well as the cultural and social aspects of our society. This has opened up a research area, namely the identification of biases introduced by recommender systems in the music context. Recent research, which focused on users of the Last.fm platform found that state-of-the-art music recommendation systems frequently tend to favor already popular items. To this end, we propose a new method for predicting music recommendations, which relies on a cluster-based quotas system. By assuming the distribution of the popularity of artists to have a latent variable that is estimated with a Gaussian Mixture, we find the underlying popularity clusters and use them to make the predictions by quotas that relate to each cluster mixing proportion, in a way that the resulting popularity distribution in the recommendations is closer to the popularity distribution seen in the data. In our experiments with the Last.fm data, our final predictions increased the recommendation frequencies of less popular artists, while preserving the specific characteristics of each algorithm. The GAP(g)$_r$ and mean average error (MAE) metrics are improved, showing that our recommendations are more accurate and approximate to the expected popularity of the artists in each user profile.