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

Abstract

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 pre-defined 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 accuracyrelated metrics to measure fairness upon decision-support metrics (e.g., precision, Hamed Zamani is currently affiliated with Microsoft.

Feature-blind fairness in collaborative filtering recommender systems

“…As a consequence, there is no such notion of groups. MADR (Mean Average Difference-Ranking) and GCE (General Cross-Entropy) [9] are also measures applicable to situations where a sensitive attribute is defined, and recommendation results are evaluated according to it. These are also not applicable in our context.…”

Unintended Bias in Language Model-driven Conversational Recommendation

“…Studies also show empirically that popular recommendation algorithms work better for males since many datasets are male-user-dominated [13]. One way to measure gender and age fairness of different recommendation models is based on generalized cross entropy (GCE) [10,11]; specifically, this work shows that a simple popularity-based algorithm provides better recommendations to male users and younger users, while on the opposite side, uniform random recommendations and collaborative filtering algorithms provide better recommendations to female users and older users [11]. Lin et al [29] study how different recommendation algorithms change the preferences for specific item categories (e.g., Action vs.…”

FaiREO: User Group Fairness for Equality of Opportunity in Course Recommendation

“…According to their definition of pairwise fairness, assuming that two items have received the same user engagement, then both protected groups should have the same likelihood of a clicked item being ranked above another relevant unclicked item. Deldjoo et al [26] proposed a generalized cross entropy measure of fairness that was based on a fair distribution of a model's performance over items or users. Yao et al [27] propose fairness metrics so that the error is fairly distributed across users.…”