Collaborative Research: CIF: Small: Interpretable Fair Machine Learning: Frameworks, Robustness, and Scalable Algorithms

机器学习算法正在彻底改变现代决策过程，从决定工作机会、评估贷款、确定大学招生到提出医疗干预措施。然而，尽管机器学习算法最近在解决大规模问题方面取得了成功，但也有人提出了严重的担忧，即它们并不完全客观，可能会无意中放大人类的偏见。拟议的研究项目通过开发可扩展的数据驱动的方法和算法来解决这一根本缺陷，这些方法和算法生成旨在实现可证明的公平性保证的可解释策略。该项目将向政策制定者或决策者通报可能的结果以及机器学习结果和社会公平/公平之间的权衡。此外，研究结果将提供指导方针，以支持在许多相关应用领域促进多样性和公平性的政策和法规。该研究充分利用了离散和稳健优化领域的最新进展，旨在寻找一种解决方法，这些方法能够忠实地解决带有公平度量的精确学习模型，提供强大的样本外公平保证，对数据集中的偏差和噪声异常值具有鲁棒性，并且可以有效地解决大规模问题实例。更具体地说，拟议的研究旨在通过子数据选择为公平学习制定有效的新框架，这些框架可以利用过去的努力，并增强学习结果的公平性。稳健的解决方案将被仔细设计，以显着减轻基于经验的方法的严重过拟合效应，并提高样本外性能。此外，还将致力于解决多阶段决策和资源分配问题中的算法公平性问题。该奖项反映了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 are revolutionizing modern decision-making processes, from deciding job offers, evaluating loans, and determining university enrollments to proposing medical interventions. However, despite the recent success of machine-learning algorithms in solving large-scale problems, serious concerns have been raised that they are not entirely objective and can inadvertently amplify human biases. The proposed research project addresses this fundamental shortcoming by developing scalable data-driven methods and algorithms that generate interpretable policies aiming for provable fairness guarantees. The project will inform the policy-makers or decision-makers about possible outcomes and tradeoffs between machine learning outcomes and social equity/fairness. Furthermore, the research results will provide guidelines to support policies as well as regulations to promote diversity and fairness in many relevant domains of application. The proposed research leverages recent advances in discrete and robust optimization, aiming for solution methodologies that faithfully address the exact learning models with fairness measures, provide strong out-of-sample fairness guarantees, are robust against bias and noisy outliers in the dataset, and can be solved efficiently for large-scale problem instances. More specifically, the proposed research aims to develop effective new frameworks for fair learning via sub-data selection that can leverage past efforts and enhance the fairness in the learning outcomes. Robust solution schemes will be carefully designed to significantly mitigate the severe overfitting effects of empirical-based methods and improve out-of-sample performance. Efforts will also be devoted to addressing algorithmic fairness in multi-stage decision-making and resource-allocation problems.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.