Exploring Social Choice Mechanisms for Recommendation Fairness in SCRUF

TL;DR

SCRUF-D tackles fairness in recommendations by treating fairness concerns as agents in a dynamic two-phase social choice framework, where each agent maps the historical recommendation outcome $L_{t-1}$ to a fairness value in $[0,1]$ such that $0$ is worst and $1$ is the target. The system online allocates agents to user arrivals and then aggregates the items' rankings using a voting rule that combines agent preferences with the base recommender. The authors compare three allocation strategies—Least Fair, Lottery, Weighted—and four voting rules—Rescoring, Borda, Copeland, RankedPairs—on synthetic data and a Microlending dataset, observing distinct fairness/accuracy frontiers and data-dependent dynamics. They provide an open-source SCRUF-D implementation in Python and demonstrate how a multi-agent social-choice approach can flexibly accommodate multiple fairness objectives and evolving user populations. The results suggest that mechanism choice and dataset characteristics jointly shape the fairness-utility tradeoff, supporting further exploration of multiple fairness definitions and dynamic balancing in recommender systems.

Abstract

Fairness problems in recommender systems often have a complexity in practice that is not adequately captured in simplified research formulations. A social choice formulation of the fairness problem, operating within a multi-agent architecture of fairness concerns, offers a flexible and multi-aspect alternative to fairness-aware recommendation approaches. Leveraging social choice allows for increased generality and the possibility of tapping into well-studied social choice algorithms for resolving the tension between multiple, competing fairness concerns. This paper explores a range of options for choice mechanisms in multi-aspect fairness applications using both real and synthetic data and shows that different classes of choice and allocation mechanisms yield different but consistent fairness / accuracy tradeoffs. We also show that a multi-agent formulation offers flexibility in adapting to user population dynamics.

Figures (5)

• Figure 1: Overview of the SCRUF architecture
• Figure 2: Accuracy vs average normalized fairness. Fairness target is at 1.0; baseline accuracy is shown in the dashed line.
• Figure 3: Fairness metric distribution for each agent (Synthetic Data)
• Figure 4: Accuracy vs average normalized fairness in segmented experiment. Fairness target is at 1.0; baseline accuracy is shown in the dashed line.
• Figure 5: Cumulative allocation of fairness agents with different allocation mechanisms