CAREER: Advancing Fair Data Mining via New Robust and Explainable Algorithms and Human-Centered Approaches

职业：通过新的稳健且可解释的算法和以人为本的方法推进公平数据挖掘

• 批准号: 2146091
• 负责人: Xiaoqian Wang
• 金额: $ 57.64万
• 依托单位: Purdue University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-08-15 至 2027-07-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2146091&HistoricalAwards=false
• 关键词: CAREER; Advancing; Fair; Data; Mining

Predictive discrimination is widespread in artificial intelligence (AI) applications that affect human life. Automated decisions can replicate, exaggerate social inequities, and even implement and legitimize new forms of discrimination. The fairness and equity of data mining and machine learning models are becoming a growing concern in many communities, but the constraints of sensitive information in data and the complexity of models bring critical challenges to building fair AI frameworks. This project focuses on undertaking fundamental research activities to advance fairness in data mining and machine learning, and to enable efficient human-machine interaction in human-centered and wellness-focused real-world problems. This project will result in algorithms and software that facilitate broader research of fair AI technologies in high-stake application areas, such as improving healthcare diversity. The project's impacts are easing humans' effort to build, adopt, and interact with fair models. Furthermore, this project will encourage underrepresented students into cutting edge computational research and contribute to graduate and undergraduate education in multidisciplinary areas.The research objective of this project is to create fair and explainable AI and human-in-the-loop control paradigm: designing a family of fair, explainable, and robust data mining algorithms with high expressive ability, faithful explanations, and rigorous theoretical foundations. From a data equity perspective, the investigator will design effective algorithms to achieve fair predictions while being able to protect sensitive information. From an algorithm perspective, the investigator will design novel explainable and robust models with rigorous theoretical guarantees on generalization ability and Pareto efficiency. From a human-machine interaction perspective, the project will promote human-in-the-loop interventions and integrate human feedback to repair incorrect or biased predictions. This research effort combines rigorous theoretical analysis with emerging application problems, and is applicable to addressing the grand challenges that society faces in building responsible data science.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.

预测性歧视在影响人类生活的人工智能（AI）应用中广泛存在。自动化决策可以复制、夸大社会不平等，甚至实施和合法化新形式的歧视。数据挖掘和机器学习模型的公平性和公平性正在成为许多社区越来越关注的问题，但数据中敏感信息的约束和模型的复杂性给构建公平的人工智能框架带来了严峻的挑战。该项目致力于开展基础研究活动，以促进数据挖掘和机器学习的公平性，并在以人为中心和健康为重点的现实世界问题中实现高效的人机交互。该项目将产生算法和软件，促进在高风险应用领域更广泛地研究公平的人工智能技术，例如改善医疗保健多样性。该项目的影响减轻了人类构建、采用公平模型并与之互动的努力。此外，该项目将鼓励未被充分代表的学生进行前沿计算研究，并为多学科领域的研究生和本科教育做出贡献。本项目的研究目标是创建公平、可解释的人工智能和人在环控制范式：设计一系列具有高表达能力、解释忠实、理论基础严谨的公平、可解释、鲁棒的数据挖掘算法。从数据公平的角度来看，研究者将设计有效的算法来实现公平的预测，同时能够保护敏感信息。从算法的角度来看，研究者将设计新颖的可解释的和鲁棒的模型，对泛化能力和帕累托效率有严格的理论保证。从人机交互的角度来看，该项目将促进人在环干预，并整合人的反馈来修复不正确或有偏见的预测。这项研究工作结合了严谨的理论分析和新兴的应用问题，适用于解决社会在建立负责任的数据科学方面面临的重大挑战。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。

项目成果

Self-Supervised Fair Representation Learning without Demographics

• 发表时间: 2022
• 作者: Junyi Chai;Xiaoqian Wang

"Why Not Other Classes?": Towards Class-Contrastive Back-Propagation Explanations

• 发表时间: 2022
• 作者: Yipei Wang;Xiaoqian Wang

Fairness without Demographics through Knowledge Distillation

• 发表时间: 2022
• 作者: Junyi Chai;T. Jang;Xiaoqian Wang

Difficulty-based Sampling for Debiased Contrastive Representation Learning doi

用于无偏差对比表示学习的基于难度的采样 doi

• 发表时间: 2023
• 期刊: IEEE Conference on Computer Vision and Pattern Recognition
• 作者: Jang, Taeuk;Wang, Xiaoqian
• 通讯作者: Wang, Xiaoqian