可信机器学习的公平性综述  

Survey on Fairness in Trustworthy Machine Learning

作　　者：刘文炎 沈楚云 王祥丰 金博[2,3] 卢兴见 王晓玲 查宏远[2,3] 何积丰 LIU Wen-Yan;SHEN Chu-Yun;WANG Xiang-Feng;JIN Bo;LU Xing-Jian;WANG Xiao-Ling;ZHA Hong-Yuan;HE Ji-Feng(Software Engineering Institute,East China Normal University,Shanghai 200062,China;School of Computer Science and Technology,East China Normal University,Shanghai 200062,China;Institute of Trusted Artificial Intelligence,Shanghai Research Institute for Intelligent Autonomous Systems,Shanghai 200092,China)

机构地区：[1]华东师范大学软件工程学院,上海200062 [2]华东师范大学计算机科学与技术学院,上海200062 [3]上海自主智能无人系统科学中心可信人工智能研究所,上海200092

出　　处：《软件学报》2021年第5期1404-1426,共23页Journal of Software

基　　金：上海市科委创新行动计划人工智能科技支持专项(20DZ1100300,20511101100);国家自然科学基金(61972155,61672231,12071145);上海市自然科学基金(19ZR1414200);国家重点研发计划(2020AAA0107400,2018YFB2101300)。

摘　　要：人工智能在与人类生活息息相关的场景中自主决策时,正逐渐面临法律或伦理的问题或风险.可信机器学习是建立安全人工智能系统的核心技术,是人工智能领域的热门研究方向,而公平性是可信机器学习的重要考量.公平性旨在研究机器学习算法决策对个人或群体不存在因其固有或后天属性所引起的偏见或偏爱.从公平表征、公平建模和公平决策这3个角度出发,以典型案例中不公平问题及其危害为驱动,分析数据和算法中造成不公平的潜在原因,建立机器学习中的公平性抽象定义及其分类体系,进一步研究用于消除不公平的机制.可信机器学习中的公平性研究在人工智能多个领域中处于起步阶段,如计算机视觉、自然语言处理、推荐系统、多智能体系统和联邦学习等.建立具备公平决策能力的人工智能算法,是加速推广人工智能落地的必要条件,且极具理论意义和应用价值.Artificial intelligence raises legal and ethical issues or risks when used to automated decision-making in areas closely related to daily life.Trustworthy machine learning is the core technology in artificial intelligence safety.It is a trending research direction,of which fairness is an essential aspect.Fairness is the absence of any prejudice or favoritism towards an individual or a group based on their inherent or acquired characteristics which are irrelevant in the particular context of decision-making.A comprehensive and structured overview of three research contents is provided,namely,fair representation,fair modeling,and fair decision-making algorithm.The potential causes and harmful consequences of unfairness are first identified in data and algorithm processing.Then,the abstract definition and primary mechanisms for eliminating unfairness are summarized.The research on fairness is at its early stage in fields such as computer vision,natural language processing,recommender systems,multi-agent systems,and federated learning.Fairness is a prerequisite for the application of machine learning,and constructing fair algorithms has theoretical significance and practical values.

关 键 词：可信人工智能 可信机器学习 公平性 统计公平 因果公平 公平表征 公平建模 公平决策 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]