Core Concepts
The core message of this article is to propose a unified framework for fairly selecting optimal solutions of integer linear programs, in order to improve fairness and transparency in decision-making processes that use integer programming.
Abstract
The article discusses the problem of fairly selecting one of the optimal solutions of an integer linear program (ILP) when there are multiple optimal solutions. The authors propose a unified framework to control the selection probabilities of the optimal solutions, in order to improve fairness and transparency in decision-making processes that use ILPs.
The key highlights and insights are:
The authors introduce the fair integer programming problem, which aims to control the selection probabilities of the optimal solutions of an ILP in order to satisfy various fairness criteria.
They propose several general-purpose algorithms to fairly select optimal solutions, such as maximizing the Nash product or the minimum selection probability, or using a random ordering of the agents as a selection criterion (Random Serial Dictatorship).
The authors establish connections between the fair integer programming problem and the literature on probabilistic social choice and cooperative bargaining, which enables them to study axiomatic properties of the proposed methods and to extend the framework to settings with cardinal preferences.
The authors evaluate the proposed methods on two specific applications, one with dichotomous and one with cardinal preferences. They find that the performance of the methods is rather application-specific, but that the Random Serial Dictatorship rule performs reasonably well on the evaluated welfare criteria in solution times similar to finding a single optimal solution.