CRII: RI: Bayesian Models for Fairness, and Fairness for Bayesian Models

在我们这个相互关联的社会中，人工智能（AI）和机器学习（ML）系统已经无处不在。每一天，机器学习系统都在影响着我们的购买决策，我们在虚拟和物理空间中的导航，我们建立的友谊，甚至我们形成的浪漫关系。这些系统自动做出的决定对现实世界有着越来越重要的影响，从信用评分到大学录取，再到刑事司法系统中再犯罪行为的预测，在美国各地，这些决定已经被用于保释和量刑决定。随着人工智能和机器学习技术对我们社会的影响越来越大，以及它们对美国的经济竞争力和技术领先地位的重要性，我们必须确保这些系统以公平和值得信赖的方式运行。最近的研究表明，数据驱动的AI和ML系统在某些情况下可能会表现出不公平和不公正的行为，例如由于隐藏在输入数据中的偏见，或者由于有缺陷的工程决策。该项目开发了一套工具，用于建模、测量和纠正人工智能和机器学习系统中的不公平和歧视行为。该研究的重点是同时解决可能发生在几个重叠维度上的算法歧视，包括性别、种族、国籍、性取向、残疾状况和社会经济阶层。本项目开发的新型人工智能技术解决了在这种情况下特别出现的两个主要技术挑战：公平衡量的不确定性和数据的相关性。在确保人工智能在性别和种族等多个受保护维度上的公平性时，随着维度数量或每个维度的值数量的增加，数据稀疏性迅速成为一个挑战。这种数据稀疏性直接导致公平性测量的不确定性。该项目将利用贝叶斯推理（专门处理不确定性的统计学分支）来管理这个问题。受保护（和其他）属性之间的相关性将使用概率图形模型（一类编码依赖关系的机器学习模型）加以利用。使用一种新的贝叶斯公平定义作为统一框架，该项目的贡献由三个相互依存的轨道组成。第一个轨道将侧重于开发通用建模技术，用于统计上有效地衡量公平性，使用潜在变量模型来产生简洁的表示，以及分层建模来实现数据效率。第二个轨道发展对抗性优化算法，训练机器学习算法在数据分布不确定时尊重公平约束。在第三个轨道上，该项目将开发确保贝叶斯推理公平性的方法，这可以用来防止推理反映负面刻板印象。这些方法将通过案例研究进行验证，这些案例研究涉及广泛的数据制度，包括人口普查收入数据建模、刑事司法累犯预测和社交媒体分析。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。

In our interconnected society, artificial intelligence (AI) and machine learning (ML) systems have become ubiquitous. Every day, machine learning systems influence our purchasing decisions, our navigation through virtual and physical spaces, the friendships we make, and even the romantic relationships we form. The decisions automated by these systems have increasingly important real-world consequences, from credit scoring, to college admissions, to the prediction of re-offending behavior in the criminal justice system, as is already being used for bail and sentencing decisions across the United States of America. With the growing impact of artificial intelligence and machine learning technologies on our society, and their importance to the economic competitiveness and technological leadership of the United States, it is imperative that we ensure that these systems behave in a fair and trustworthy manner. Recent studies have shown that data-driven AI and ML systems can in some cases exhibit unfair and unjust behavior, for example due to biases hidden in the input data, or because of flawed engineering decisions. This project develops a suite of tools for modeling, measuring, and correcting unfair and discriminatory behavior in AI and ML systems. The research focuses on simultaneously addressing algorithmic discrimination that may occur across several overlapping dimensions, including gender, race, national origin, sexual orientation, disability status, and socioeconomic class. The novel AI techniques developed in this project address the two main technical challenges which specifically arise in this context: uncertainty in the measurement of fairness, and correlations in the data.When ensuring AI fairness regarding multiple protected dimensions such as gender and race, data sparsity rapidly becomes a challenge as the number of dimensions, or the number of values per dimension, increase. This data sparsity directly results in uncertainty in the measurement of fairness. The project will leverage Bayesian inference, a branch of statistics which specifically addresses uncertainty, to manage this issue. Correlations between the protected (and other) attributes will be leveraged using probabilistic graphical models, a class of machine learning models which encode dependence relationships. Using a novel Bayesian definition of fairness as a unifying framework, the project's contributions consist of three interdependent tracks. The first track will focus on developing general modeling techniques for the statistically efficient measurement of fairness, using latent variable models to produce parsimonious representations, and hierarchical modeling to achieve data efficiency. The second track develops adversarial optimization algorithms to train machine learning algorithms to respect fairness constraints when the data distribution is uncertain. In the third track, the project will develop methods for ensuring fairness in Bayesian inference, which can be used to prevent the inferences from reflecting negative stereotypes. The methods will be validated with case studies on applications across a wide range of data regimes, including modeling census income data, criminal justice recidivism prediction, and social media analytics.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.