EAGER: Causal Bayesian Network-Based Discrimination Discovery and Prevention

围绕收集和使用客户数据来做出就业、信贷和保险等重要决策，已经建立了各种商业模式。人们越来越担心歧视，因为数据分析技术可能被用来根据个人的人口统计信息，如性别、年龄、婚姻状况、种族、宗教或信仰、少数民族成员身份、残疾或疾病来不公平地对待他们。必须开发预测性决策模型，以便输入其中的数据和在它们的帮助下作出的决定不会受到歧视。这项急切的研究设计了实用的技术，以准确地检测和消除用于建立决策模型的数据集中的歧视。这项研究的一个主要成果是一个统一的框架和一个发现和消除歧视的原型系统。该系统可以帮助来自弱势群体的个人确定他们是否得到公平的对待，并帮助来自组织的决策者确保他们的预测决策模型是无歧视的。现有的歧视发现方法主要基于相关性或关联性，不能准确地发现真实的歧视。此外，每一项都只针对一到两种类型的歧视。这项研究根据歧视是发生在整个系统、发生在一个子系统还是发生在一个人身上，以及歧视是对决策的直接影响还是间接影响来对歧视进行分类。然后，本研究开发了一个统一的基于因果贝叶斯网络的框架，该框架考虑了歧视和一般因果之间的区别，并通过受保护属性和决策之间的不同路径将直接歧视和间接歧视建模为因果效应。它可以准确地捕捉和测量系统、群体和个人层面上的各种歧视。然后，该研究开发了新的歧视发现和预防模型和算法。该研究还建立了基于各种度量的不同类型歧视的模拟和评估方法的测试框架，并将歧视发现和预防算法集成到开源数据挖掘和机器学习软件系统中。

Various business models have been built around the collection and use of customer data to make important decisions like employment, credit, and insurance. There are increasing worries of discrimination as data analytics technologies could be used to unfairly treat individuals based on their demographic information such as gender, age, marital status, race, religion or belief, membership in a national minority, disability, or illness. It is imperative to develop predictive decision models, such that the data that goes into them and the decisions made with their assistance are not subject to discrimination. This EAGER research designs practical techniques to accurately detect and remove discrimination from the datasets used to build decision models. A primary outcome of this research is a unifying framework and a prototype system for discrimination discovery and removal. This system can help individuals from disadvantaged groups determine whether they are fairly treated and help decision makers from organizations ensure their predictive decision models are discrimination free. Existing discrimination discovery approaches are mainly based on correlation or association and cannot accurately discover the true discrimination. In addition, each of them targets on one or two types of discrimination only. This research categorizes discrimination based on whether discrimination is across the whole system, occurs in one subsystem, or happens to one individual, and whether discrimination is a direct effect or an indirect effect on the decision. This research then develops a unifying causal Bayesian network based framework that takes into consideration the distinctions between discrimination and general causalities and models both direct discrimination and indirect discrimination as causal effects via different paths between protected attributes and the decision. It can accurately capture and measure various types of discrimination at system, group, and individual levels. The research then develops novel discrimination discovery and prevention models and algorithms. The research also builds a testing framework for simulating different types of discrimination and evaluating the approaches based on various metrics, and integrates the discrimination discovery and prevention algorithms into an open source data mining and machine learning software system.