Businesses and organizations must ensure that their algorithmic decision-making is fair in order to meet legislative, ethical, and societal demands. For example, decision-making in automated hiring must not discriminate with respect to gender or race. To achieve this, prior research has contributed approaches that ensure algorithmic fairness in machine learning predictions, while comparatively little effort has focused on algorithmic fairness in decision models, specifically off-policy learning. In this paper, we propose a novel framework for fair off-policy learning: we learn decision rules from observational data under different notions of fairness, where we explicitly assume that observational data were collected under a different -- potentially biased -- behavioral policy. For this, we first formalize different fairness notions for off-policy learning. We then propose a machine learning approach to learn optimal policies under these fairness notions. Specifically, we reformulate the fairness notions into unconstrained learning objectives that can be estimated from finite samples. Here, we leverage machine learning to minimize the objective constrained on a fair representation of the data, so that the resulting policies satisfy our fairness notions. We further provide theoretical guarantees in form of generalization bounds for the finite-sample version of our framework. We demonstrate the effectiveness of our framework through extensive numerical experiments using both simulated and real-world data. As a result, our work enables algorithmic decision-making in a wide array of practical applications where fairness must ensured.

翻译：企业和组织必须确保其算法决策的公平性，以满足法律、伦理和社会需求。例如，自动化招聘中的决策不得因性别或种族而产生歧视。为此，现有研究提出了确保机器学习预测中算法公平性的方法，但针对决策模型（特别是离线策略学习）中算法公平性的工作相对较少。本文提出了一种新颖的公平离线策略学习框架：我们从不同公平性概念下的观测数据中学习决策规则，并明确假设这些观测数据是在一种不同的（可能带有偏见的）行为策略下收集的。为此，我们首先形式化定义了离线策略学习中的多种公平性概念，进而提出一种机器学习方法来学习满足这些公平性概念的最优策略。具体而言，我们将公平性概念重新表述为可从有限样本中估计的无约束学习目标，并利用机器学习在数据公平表示约束下最小化该目标，从而使所得策略满足我们的公平性概念。我们进一步提供了理论保证，即针对框架有限样本版本的泛化界。通过使用模拟数据和真实数据进行的大量数值实验，我们验证了框架的有效性。最终，我们的工作使得在需要确保公平性的广泛实际应用中实现算法决策成为可能。