A pessimist's approach to one-sided matching

TL;DR

This paper investigates decomposing probabilistic matchings into efficient deterministic matchings, ensuring fairness and strategy-proofness, with algorithms and bounds for specific mechanisms, supported by real-world data evaluations.

Contribution

It introduces a polynomial-time algorithm for decomposing certain probabilistic assignments into Pareto-efficient matchings, maximizing assigned agents, and analyzes bounds for other mechanisms.

Findings

Polynomial-time algorithm for Probabilistic Serial mechanism decomposition

Decomposition guarantees at least the expected number of assigned agents

Random Serial Dictatorship assigns at least half of the optimal agents

Abstract

Inspired by real-world applications such as the assignment of pupils to schools or the allocation of social housing, the one-sided matching problem studies how a set of agents can be assigned to a set of objects when the agents have preferences over the objects, but not vice versa. For fairness reasons, most mechanisms use randomness, and therefore result in a probabilistic assignment. We study the problem of decomposing these probabilistic assignments into a weighted sum of ex-post (Pareto-)efficient matchings, while maximizing the worst-case number of assigned agents. This decomposition preserves all the assignments' desirable properties, most notably strategy-proofness. For a specific class of probabilistic assignments, including the assignment by the Probabilistic Serial mechanism, we propose a polynomial-time algorithm for this problem that obtains a decomposition in which all matchings assign at least the expected number of assigned agents by the probabilistic assignment, rounded down, thus achieving the theoretically best possible guarantee. For general probabilistic assignments, the problem becomes NP-hard. For the Random Serial Dictatorship mechanism, we show that the worst-case number of assigned agents is at least half of the optimal, and that this bound is asymptotically tight. Lastly, we propose a column generation framework for the introduced problem, which we evaluate both on randomly generated data, and on real-world school choice data from the Belgian cities Antwerp and Ghent.