Collaborative Research: III: MEDIUM: Responsible Design and Validation of Algorithmic Rankers

数据驱动型系统采用算法来帮助人类在招聘和就业、学校和大学录取、信贷和贷款以及大学排名等关键领域做出判断。由于它们对个人、人口群体、机构和整个社会的影响，将公平、问责制和透明度考虑纳入这些系统的设计、验证和使用至关重要。目前该领域的研究主要集中在分类和预测任务上。然而，评分和排名也被广泛使用，并且引起了许多为分类设计的方法无法处理的问题，因为分类标签一次应用一个项目，而排名则明确地设计为比较项目。这个项目的重点是基于分数的排名算法，它根据一个“简单”的评分公式对一组候选人进行排序。这种排序器被广泛应用于关键领域，因为它们比复杂的学习模型更容易设计、理解和证明。然而，即使这些看似简单透明的排名也可能产生反直觉的结果，不公平地将属于弱势群体的候选人降级，并且由于对输入数据或评分公式的微小变化的敏感性，容易受到操纵。由于被排序的数据与排名者之间的相互作用、数据内部的复杂结构以及平衡多个目标的需要，解决这些问题是具有挑战性的。该项目考虑了算法排名的负责任设计和验证所固有的核心技术挑战，并追求三个协同目标。目标1是开发方法来量化项目属性的影响，以及关于属性表示和预处理的特定工程选择对排名结果（验证）的影响。然后，这些信息用于指导数据科学家选择与他们对质量或适当性（设计）的理解相对应的评分函数。目标2是开发方法来量化数据不确定性、评分公式的微小变化或两者对排名结果的影响（验证）。然后，这些信息用于指导数据科学家干预数据采集和预处理，以减少不确定性，并选择一个足够稳定的评分函数（设计）。目标3是考虑到多重公平目标和潜在的交叉歧视（验证），开发方法来量化来自代表性不足或历史上处于不利地位的群体的候选人在排名结果中缺乏公平性。然后使用这些信息来确定可行的权衡，并帮助数据科学家在这些权衡中制定增强公平性的干预措施（设计）。这项工作的结果将影响关键领域的评分和排名，如教育项目招生、招聘和大学排名。从这项工作中获得的见解将有助于在适当的时候进行技术干预，并确定技术干预不足的情况，以及应收集更多数据或应使用替代筛选过程的情况。该项目还将包括教学和指导、公共教育和推广，以及扩大代表性不足群体成员在计算机领域的参与。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Data-driven systems employ algorithms to aid human judgment in critical domains like hiring and employment, school and college admissions, credit and lending, and college ranking. Because of their impacts on individuals, population groups, institutions, and society at large, it is critical to incorporate fairness, accountability, and transparency considerations into the design, validation, and use of these systems. Current research in this area has mainly focused on classification and prediction tasks. However, scoring and ranking are also used widely, and raise many concerns that methods designed for classification cannot handle because classification labels are applied one item at a time, whereas ranking is explicitly designed to compare items. This project is focused on algorithmic score-based rankers that sort a set of candidates based on a “simple” scoring formula. Such rankers are widely used in critical domains because of the premise that they are easier to design, understand, and justify than complex learned models. Yet, even these seemingly simple and transparent rankers may produce counter-intuitive results, unfairly demote candidates that belong to disadvantaged groups, and be prone to manipulation due to sensitivity to slight changes in the input data or in the scoring formula. Addressing these issues is challenging due to the interplay between the data being ranked and the ranker, the complex structure within the data, and the need to balance multiple objectives.This project considers the core technical challenges inherent in the responsible design and validation of algorithmic rankers, and pursues three synergistic aims. Aim 1 is to develop methods to quantify the impact of item attributes, and of specific engineering choices regarding attribute representation and pre-processing, on the ranked outcome (validation). This information is then used to guide the data scientist in selecting a scoring function that corresponds to their understanding of quality or appropriateness (design). Aim 2 is to develop methods to quantify the impact of data uncertainty, of slight changes in the scoring formula, or both, on the ranked outcome (validation). This information is then used to guide the data scientist in intervening on data acquisition and pre-processing to reduce uncertainty, and in selecting a scoring function that is sufficiently stable (design). Aim 3 is to develop methods to quantify lack of fairness in ranked outcomes, with respect to candidates from under-represented or historically disadvantaged groups, in view of multiple fairness objectives and potential intersectional discrimination (validation). This information is then used to identify feasible trade-offs and assist the data scientist in navigating these trade-offs to enact fairness-enhancing interventions (design). Outcomes of this work will impact the practice of scoring and ranking in critical domains like educational program admissions, hiring, and college ranking. Insights from this work will enable technical interventions when appropriate, and also identify cases where they are insufficient, and where more data should be collected or an alternative screening process should be used. This project will also include teaching and mentoring, public education and outreach, and broadening participation of members of under-represented groups in computing.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.