A flexible framework for evaluating user and item fairness in recommender systems

One common characteristic of research works focused on fairness evaluation (in machine learning) is that they call for some form of parity (equality) either in treatment—meaning they ignore the information about users’ memberships in protected classes during training—or in impact—by enforcing proportional beneficial outcomes to users in different protected classes. In the recommender systems community, fairness has been studied with respect to both users’ and items’ memberships in protected classes defined by some sensitive attributes (e.g., gender or race for users, revenue in a multi-stakeholder setting for items). Again here, the concept has been commonly interpreted as some form of equality—i.e., the degree to which the system is meeting the information needs of all its users in an equal sense. In this work, we propose a probabilistic framework based on generalized cross entropy (GCE) to measure fairness of a given recommendation model. The framework comes with a suite of advantages: first, it allows the system designer to define and measure fairness for both users and items and can be applied to any classification task; second, it can incorporate various notions of fairness as it does not rely on specific and predefined probability distributions and they can be defined at design time; finally, in its design it uses a gain factor, which can be flexibly defined to contemplate different accuracy-related metrics to measure fairness upon decision-support metrics (e.g., precision, recall) or rank-based measures (e.g., NDCG, MAP). An experimental evaluation on four real-world datasets shows the nuances captured by our proposed metric regarding fairness on different user and item attributes, where nearest-neighbor recommenders tend to obtain good results under equality constraints. We observed that when the users are clustered based on both their interaction with the system and other sensitive attributes, such as age or gender, algorithms with similar performance values get different behaviors with respect to user fairness due to the different way they process data for each user cluster.