CAREER: Foundations for Fair Social Network Analysis

越来越多的决策受到算法的影响，虽然这给社会带来了明显的好处，但可能的危害也开始变得明显。在这种危害的一个显著例子中，最近的研究表明，算法的使用可能会延续或创造新的不道德歧视形式。在社会网络分析领域，个体通过一组复杂的连接与他人互动，其中包含许多令人担忧的应用。例如，一些信用评分公司使用社交网络数据（例如，朋友和家庭关系）为个人分配信用评分；在这种应用中，重要的是要确保在这种过程中使用的算法不会因种族或性别等受保护属性而无意中歧视个人。在这个项目中，研究者将开发公平社会网络分析领域的算法。公平社会网络分析的目标是了解网络结构和网络算法如何导致对个人群体的系统性伤害，并针对此类情况提出补救措施。该项目是公平社会网络分析领域的首批项目之一，它的贡献将对从事网络数据工作的从业者和研究人员都有价值。由此产生的算法可用于在线广告定位和社交媒体好友推荐等应用。除了该项目的科学目标之外，研究者还将开展与伦理算法设计领域的课程开发相关的活动，共同开发针对律师的继续法律教育研讨会（重点是算法伦理），向当地农村学生提供服务，并为与社区组织合作的技术人员编写指南手册。虽然关于机器学习公平性的研究越来越多，但现有的方法没有考虑点之间的依赖关系，因此不适用于链接预测或社区检测等网络任务。该项目将包含三个任务。在第一个任务中，研究者将设计测试来确定网络结构或网络分析中是否存在不公平。在第二项任务中，研究者将设计算法来减少网络分析中的不公平性。在第三个任务中，研究者将设计修改网络的算法，以减少其结构中的不公平。本项目的主要贡献将是：(1)开发网络公平的正式定义，(2)创建算法测试以检测网络结构中的不公平，(3)设计测试以确定网络分析算法对特定属性的依赖，(4)开发算法以减轻这种依赖，以及(5)开发算法以修改网络以减少错误偏差。教育和推广活动包括算法设计课程的开发，以及与法律界和社区组织的接触。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。

An increasing amount of decision making is influenced by algorithms, and while this has resulted in clear benefits to society, the possible harms are starting to become apparent. In a striking example of such harms, recent works have argued that use of algorithms can perpetuate or create new forms of unethical discrimination. The area of social network analysis, in which individuals interact with one another through a complex set of connections, contains many worrisome applications. For example, some credit scoring companies use social network data (e.g., friends and family connections) to assign credit scores to individuals; and in such applications, it is important to ensure that the algorithms used in such a process are not inadvertently discriminating against individuals on the basis of protected attributes like race or sex. In this project, the investigator will develop algorithms for the area of fair social network analysis. The goal of fair social network analysis is to understand how network structure and network algorithms may lead to systematic harm against groups of individuals, and to propose remedies for such cases. This project is among the first in the area of fair social network analysis, and its contributions will be of value to both practitioners and researchers working with network data. The resulting algorithms may be used in applications like online advertisement targeting and social media friendship recommendation. In addition to the scientific objectives of the project, the investigator will conduct activities related to course development in the area of ethical algorithm design, co-development of Continuing Legal Education seminars for attorneys (focusing on ethics of algorithms), outreach to local rural students, and development of a handbook on guidelines for technologists working with community organizations.While there is a growing body of research on fairness in machine learning, existing methods do not consider dependencies between points, and so do not apply to network tasks like link prediction or community detection. The project will contain three tasks. In the first task, the investigator will design tests for determining whether unfairness exists in a network structure or network analysis. In the second task, the investigator will design algorithms to reduce the unfairness in network analysis. In the third task, the investigator will design algorithms for modifying a network to reduce unfairness in its structure. The main contributions of this project will be (1) Development of formal definitions of fairness in networks, (2) Creation of algorithmic tests to detect unfairness in network structure, (3) Design of tests to determine reliance of a network analysis algorithm on a particular attribute, (4) Development of algorithms to mitigate such reliance, and (5) Development of algorithms to modify a network to reduce wrongful bias. Education and outreach activities include course development in algorithmic design, and engagement with the legal community and community organizations.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.