公平机器学习:概念、分析与设计

Fair Machine Learning:Concepts,Analysis,and Design

随着人工智能的发展,机器学习技术越来越多地应用于社会各个领域,用以辅助或代替人们进行决策,特别是在一些具有重要影响的领域,例如,信用程度评级、学生质量评估、福利资源分配、疾病临床诊断、自然语言处理、个性信息推荐、刑事犯罪判决、无人驾驶等.如何在这些应用中确保决策公平或者无偏见?如何在这些应用中保护弱势群体的利益?这些问题直接影响到社会和公众对机器学习的信任,影响到人工智能技术的应用与系统的部署.通过系统梳理和全面剖析近年来的工作,对机器学习公平性或公平机器学习的定义及度量进行了解释及对比;从机器学习的全生命周期出发,对不同环节中出现的各类偏见及其发现技术进行了归类及阐释;从预处理、中间处理和后处理三个阶段,对公平机器学习的设计技术进行了介绍和分析;从可信赖人工智能全局出发,对公平性与隐私保护、可解释性之间的关系、影响及协同解决方案进行了阐述;最后对公平机器学习领域中亟待解决的主要问题、挑战及进一步研究热点进行了讨论.

With the development of artificial intelligence,machine learning techniques is increasingly used in many social domains to assist or replace humankinds in decision-making,especially in some critical areas,such as,credit rating,students’qualification evaluation,welfare resource allocation,clinical diagnosis,natural language processing,personalized information recommendation,criminal judgment,autonomous vehicles and so on.Due to the intrinsic and technical characteristics of machine learning itself,its prediction and decision-making will inevitably produce a certain degree of bias or unfairness,which has gradually attracted the attention of scientific research,industry practitioners and the public.How to ensure fair or unbiased decisions in machine learning?How to protect the interests of disadvantaged groups in these applications?These issues have important impacts on the society and the public’s confidence in machine learning and affect the application of artificial intelligence technology and the deployment of artificial intelligence systems.Fairness has been one of the basic supporting capabilities of trustworthy artificial intelligence,and machine learning with fairness is referred to as fair machine learning.In this paper,the concepts of fairness,the methods of discovering unfair or biased discrimination and the design techniques of fair machine learning are reviewed and discussed.The detailed contents include the followings.Firstly,discrimination and bias are terminologies related to unfairness,and unfair behavior is known as biased behavior or discriminatory behavior.Since the taxonomy of discrimination and biases is helpful to understand and evaluate the fairness,direct discrimination,indirect discrimination,interpretable discrimination,uninterpretable discrimination,statistical discrimination and systematic discrimination are explained.In the framework of statistics,similarity and causal inference,the definitions and quantification of fairness in machine learning are categorized and explained.Secondly,the bias or prejudice is the main source of discrimination and unfairness.The training data and algorithms involved in machine learning can have biases that lead to unfair model predictions.From the perspectives of data,algorithm and human-computer interaction,the biases in the life cycle of machine learning are classified and discussed.The techniques to discover biases in machine learning,such as association rule mining,k-nearest neighbor classification,probabilistic causal network,and privacy attack and deep learning methods,are illustrated.Meanwhile,the design methodologies of fair machine learning have been undertaken roughly in three directions.On the view of specific applicable tasks,fair natural language processing,fair face recognition,fair recommendation system,fair classification,fair regression and fair clustering are elaborated.In light of particular machine learning algorithms,fair representation and fair adversarial learning are discoursed.From the life cycle of machine learning,preprocessing methods,intermediate processing methods and post-processing methods are expounded.Then,for the trustworthy artificial intelligence,the recent studies regarding anonymous protection,secure multi-party computing and security attack and defense for fair machine learning are promising works,which are briefly introduced.The explainability can help to discover algorithmic bias in machine learning models,on which some preliminary attempts are conducted,also being described.Finally,the main problems,challenges and hot topics in the research of fair machine learning,such as evaluation and testing of fair machine learning,novel modes of fair machine learning and ethically aligned machine learning,are presented.