CIF: Small: Foundations and Applications of Blind Subgroup Robustness

机器学习算法可能会在某些子组中呈现歧视性行为，这意味着总体人口的一部分明显得不到该模型的服务，从而使决策不公平。解决这一挑战的最常见方法考虑了算法在训练期间可以访问一组预定义的受保护的子组，并且目标是学习满足这些子组中的公平/稳健性的特定概念的模型。一般来说，只有降低受益群体的表现，而不一定改善弱势群体和受保护群体的表现，才能实现完美的公平。这与无害公平的道德和法律概念相冲突，这些概念适用于服务质量至高无上的情况，例如在健康领域。为了解决这个问题，这项工作考虑了公平性和子群健壮性的概念，以保证不会对任何子群造成不必要的伤害。该项目超越了这一点，因为它考虑了子组或人口统计数据先验未知的情况，甚至可能会随着时间和算法部署而变化。该项目将盲子组健壮性和公平性这些概念引入向后兼容性领域，目标是保证新的机器学习算法与以前的算法兼容；以及联合学习领域，其中多个站点为了互利而共享数据。最后，研究了所提出的盲子群稳健性和无不必要伤害子群稳健性与因果推理之间的潜在联系。该项目首先正式研究了盲子群和无不必要伤害(Pareto最优)子群的稳健性，其中机器学习算法需要对数据的所有可能的子群具有健壮性(给定最小子群大小)，而不必事先知道子群的定义特征。这是形式化的研究，包括权衡和成本保护未知子组和相应的优化算法；还包括数据和优化不确定性的概念，以模型化一个子组可能做出的潜在牺牲，为其他人的利益。这种盲子群稳健性的形式化研究是机器学习领域的一个新兴领域，该项目提供了一个基本和统一的观点，将理论与实践相结合，并为决策者提供关键信息。然后，该项目将工作扩展到向后兼容领域，目标是使所有潜在的子组同样向后兼容；以及联合学习，其中子组的公平性和稳健性既考虑到竖井/参与者之间，也考虑到每个竖井本身。最后，由于不变特征和因果关系之间的密切数学联系，该项目进一步考虑了提出的盲子群稳健性的统一框架来研究自动发现的临界子群之间的联系、它们的特征和因果关系。健康应用程序为这里开发的框架提供了一个独特的试验台。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Machine-learning algorithms may present discriminatory behavior across certain subgroups, meaning that segments of the overall population are measurably under-served by the model, rendering the decisions unfair. The most common approaches to address this challenge consider that the algorithm has access to a set of predefined protected subgroups during training, and the goal is to learn a model that satisfies a certain notion of fairness/robustness across these subgroups. Perfect fairness can, in general, only be achieved by degrading the performance of the benefited subgroups without necessarily improving the disadvantaged and protected ones. This conflicts with ethical and legal notions of no-harm fairness, which are appropriate where quality of service is paramount, for example in health. To address this, this work considers notions of fairness and subgroup robustness that guarantee no unnecessary harm is done to any subgroup. The project goes beyond this since it considers the case where the subgroups or demographics are not known a priori and might even change with time and algorithm deployment. The project brings these concepts of blind and no-harm subgroup robustness and fairness to the area of backwards compatibility, where the goal is to guarantee that new machine-learning algorithms are compatible with previous ones; and to the area of federated learning, where multiple sites share data for the sake of mutual benefit. Lastly, potential connections of the proposed blind and no unnecessary-harm subgroup robustness with causal inference are investigated. The project first formally studies blind and no-unnecessary-harm (Pareto optimal) subgroup robustness, where the machine-learning algorithm needs to be robust to all possible subgroups of the data (given a minimal subgroup size), without necessarily knowing in advance the subgroups' defining characteristics. This is formally studied, including the tradeoffs and costs of protecting unknown subgroups and the corresponding optimization algorithm; concepts of data and optimization uncertainty are also included to model potential sacrifices a subgroup can make in benefit of others. Such formal study of blind subgroup robustness is an emerging field in the machine-learning community, and this project provides a fundamental and unifying view of it, combining theory with practice and critical information for policy makers. The project then extends the work to the area of backwards compatibility, with the goal to make all potential subgroups equally backwards compatible; and to federated learning, where the subgroup fairness and robustness is considered both across the silos/participants and inside each silo itself. Finally, thanks to the close mathematical connection between invariant features and causality, the project further considers this proposed unifying framework of blind subgroup robustness to study connections between the automatically discovered critical subgroups, their features, and causality. Health applications provide a unique testbed for the frameworks developed here.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.