III: Small: Fairness and Control of Exposure in Ranking

通过机器学习训练的排名功能在当今的在线系统中随处可见，在这些系统中，它们几乎被用来对从产品、电影到求职者的任何东西进行排名。通过决定单个商品的排名位置以及用户找到它们的难易程度，排名功能极大地影响了哪些产品被购买，哪些候选人获得工作，以及哪些电影被流媒体播放。这就提出了公平的问题，声称排名功能应该对系统的用户和被排名的项目都是公平的。该项目为排名函数开发了新的公平标准，以及设计和学习具有公平性保证的排名函数的新方法。该项目基于一个排名系统作为双边市场的模型，其中效用不仅流向发出查询的用户，也流向被排名的项目。遗憾的是，目前使用的几乎所有学习排序(LTR)方法都只优化用户的平均效用，这可能导致对项目和少数用户组的不公平对待。为了克服这一缺陷，该项目开发了能够强制实施所需的公平约束的LTR方法。这些新方法可以弥补对用户群体的不同待遇(例如，招聘系统中性别偏见的放大)、在线市场的市场集中度以及参与在线系统的动态(例如，两极分化)。为了实现这些目标，该项目解决了不公平的内生和外生原因。外生原因是由于训练数据中的偏差，这往往导致致富-致富的动力。然而，即使在训练时使用无偏数据，内生原因在设计LTR算法时也会导致不公平。因此，项目中开发的LTR方法解决了内生和外生原因。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。

Ranking functions trained via machine learning are ubiquitous in today’s online systems, where they are used to rank virtually anything - from products and movies to job candidates. By deciding where individual items get ranked and how easily users can find them, the ranking function greatly influences which products get purchased, which candidates get a job, and which movies get streamed. This raises questions of fairness, asserting that the ranking functions should be fair to both the users of the systems as well as to the items being ranked. The project develops new fairness criteria for ranking functions, as well as new methods for designing and learning ranking functions with fairness guarantees.The project is based on a model of ranking systems as two-sided markets, where utility goes not only to the users issuing the queries, but also to the items that are being ranked. Unfortunately, virtually all learning-to-rank (LTR) methods in use today only optimize the average utility to the users, which can lead to unfair treatment of the items and of minority user groups. To overcome this deficiency, the project develops LTR methods that can enforce desirable fairness constraints. These new methods can remedy disparate treatment of user groups (e.g. amplification of gender bias in a hiring system), market concentration in online markets, and the dynamics of participation in online systems (e.g. polarization). To achieve these goals, the project addresses both endogenous and exogenous causes of unfairness. Exogenous causes are due to biases in the training data, which often lead to rich-get-richer dynamics. However, even when trained with unbiased data, causes endogenous in the design of the LTR algorithm can lead to unfairness. Therefore, the LTR methods developed in the project address both endogenous and exogenous causes.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.