Mechanism-Informed Learning for Fair Division

Fair division provides a simple yet powerful framework for modeling fairness in resource allocation. While existing literature typically assumes complete information about preferences, many practical scenarios involve incomplete preferences, posing challenges in computing fair allocations. In this paper, we propose mechanisminformed preference learning, a framework that integrates neural networks with differentiable approximations of classical fair division mechanisms—adjusted winner, round-robin, and movingknife—to estimate fair allocations from incomplete preferences. Experiments on real-world household chore preference data show that our mechanism-informed framework achieves fairer allocations, compared to methods without mechanism information.

Authors

• Ryota Maruo Kyoto University
• Tomohiko Yokoyama The University of Tokyo
• Ayumi Igarashi The University of Tokyo & RIKEN AIP
• Koh Takeuchi Kyoto University & RIKEN AIP

Keywords

• Fair Division
• Machine Learning