III:Small: Counterfactually Fair Machine Learning through Causal Modeling

机器学习模型在现代社会中扮演着越来越重要的角色。围绕历史数据的收集和使用，已经建立了各种机器学习模型，以自动做出就业、贷款审批、保险、保释和刑事判决等重要决策。必须确保在机器学习模型的帮助下做出的决策不受歧视和社会偏见的影响。为了实现这一目标，必须回答一些基本问题，包括如何定义公平标准以及如何将其纳入机器学习算法以构建公平模型。该项目旨在为应对这些挑战建立一个统一的框架。该框架建立在因果推理理论的基础上，以回答在判断歧视和公平方面至关重要的“如果”问题。该项目的成功结果将导致对因果关系应用于公平的理论理解的进步，并有助于公平机器学习中有限的知识基础。该教育计划将通过课程和论文项目，让本科生和研究生提高他们在机器学习和人工智能领域解决问题的知识和技能，并吸引高中生，特别是那些来自代表性不足群体的高中生从事STEM领域的职业。本项目提出了一个统一的公平机器学习框架，该框架涵盖了不同类型的公平措施，如完全歧视、直接/间接歧视或反事实公平，不同水平的公平，如群体水平或个人水平，混合数据类型，如分类或数值，因果模型中不同形式的方程，以及数据和模型的公平性。在该框架内，研究人员将通过挖掘不可识别性的深层原因和推导不可识别因果效应的数学界限来解决因果推理中的不可识别性挑战。研究人员将利用开发的技术和工具，对现有的公平概念进行量化度量，并提出与该框架兼容的新公平概念。研究人员将把提出的公平措施纳入机器学习模型构建中。最后，研究人员将通过深度生成模型处理混合类型变量，并将马尔可夫假设放宽为半马尔可夫模型，使所提出的框架更普遍适用。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。

Machine learning models play increasingly important roles in modern society. Various machine learning models have been built around the collection and use of historical data to automatically make important decisions like employment, loan approval, insurance, bail and criminal sentencing. It is imperative to ensure that the decisions made with the assistance of machine learning models are not subject to discrimination and social biases. To achieve this goal, fundamental questions including how to define fairness criteria and how to incorporate them into machine learning algorithms to build fair models must be answered. This project is to establish a unified framework for addressing these challenges. The framework is built upon causal inference theories to answer “what if” questions that are critical in judging discrimination and fairness. The successful outcome of this project will lead to advances in theoretical understanding of the application of causality to fairness and contribute to the limited base of knowledge in fair machine learning. The education program will involve, through courses and thesis projects, undergraduate and graduate students to enhance their knowledge and skills in solving problems in machine learning and artificial intelligence, and attract high school students especially those from underrepresented groups to pursue careers in STEM.This project proposes a unified framework for fair machine learning, which covers different types of fairness measures such as total discrimination, direct/indirect discrimination or counterfactual fairness, different levels of fairness such as group-level or individual-level, mixed data type such as categorical or numerical, different forms of equations in the causal model, and fairness in both the data and the model. Within the framework, the investigators will address the unidentifiability challenge in causal inference by digging into the deep-seated reasons of unidentifiability and deriving mathematical bounds for unidentifiable causal effects. With the developed techniques and tools, the investigators will develop quantitative measure for existing fairness notions as well as propose new fairness notions that are compatible with the framework. The investigators will incorporate proposed fairness measures into machine learning model construction. Finally, the investigators will make the proposed framework more generally applicable by handling the mixed-type variables via deep generative models and relaxing the Markovian assumption as semi-Markovian models.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.