Analysis of Biases in Calibrated Recommendations

Carlos Rojas, David Contreras, Maria Salamó

Abstract: While recommender systems have mainly focused on the effectiveness of their results, beyond-accuracy perspectives have been recently explored. One of the most prominent is algorithmic bias, which analyzes if existing imbalances in the input data are exacerbated in the produced recommendations. On the other hand, calibrated recommendations ensure that the recommendations reflect the distribution of the original preferences of each user (e.g., in terms of item genres). In this paper, we connect these two perspectives, to analyze how the original calibration method deals with the bias in the state-of-the-art recommendation models. Our analysis on real-world data shows that the calibration effectiveness is impacted by how a recommendation model handles bias.