Causality, Counterfactuals and Meta-learning to Address the Complexity of Fairness in Data Science and Machine Learning

到目前为止，解决机器学习公平性问题的方法大多无法解决社会中系统性不平等的复杂性。朝着解决这一问题的正确方向迈出的一步是创建旨在确定和消除根本原因的统计方法。我认为这可以通过更具体、更复杂的因果和反事实模型来实现，以推断多种原因、结构，并避免对社会类别进行假设。我建议将其扩展到元学习的可能性，以进一步理解不平等的结构，这也可以作为政策制定和审计的技术基础。机器学习在准确预测结果方面非常有效，从而为快速有效地分配稀缺的社会资源提供了机会。因此，机器学习已迅速在社会技术系统的高风险决策中获得存在，这些系统涉及人类，机器和社会之间的复杂互动。随着机器学习在这一领域的发展，它在刑事司法系统、医疗保健和教育系统中的存在表明，这些算法很容易复制和夸大世界上存在的歧视，造成重大伤害。这导致了机器学习中一个新领域——公平的发展，并创建了诸如公平、问责制和透明度（FAccT）以及算法、机制和优化中的公平和访问（EAAMO）等会议。因此，近年来，机器学习研究人员已经做出了巨大的努力，以减轻其技术加剧的偏见和歧视，然而，在当前的研究领域存在重大的差距和失败。公平算法在特定环境之外是不可推广的，社会不平等是持久的、系统性的和复杂的，这在技术工作中没有反映出来。除了对偏见的基本先验论证之外，几乎没有人做过什么来整合社会科学，大多数公平工作都是在质量保证方面的权宜之计，而不是试图找到问题的根源。本研究计划包括关于如何将社会不平等的复杂性（如交叉理论和次边缘性）整合到统计和机器学习中的问题和想法，以实现公平的机器学习版本，正确地与社会的复杂性一起工作。虽然我的研究问题是由该领域以前的大量工作形成的，但我的具体问题主要是基于旨在提高公平因果模型复杂性的工作，例如公平排名和影响补救的交叉性。在我的博士工作中，我想解决的主要问题是：能否通过特定的因果和反事实模型来开发和理解机器学习中更复杂的公平概念，以推断多种原因和结构？这能否与元学习相结合，学习几种算法，从而学习社会中几种不同的歧视层次？

The complexity of systemic inequalities in society has mostly so far eluded the methods taken to address fairness in machine learning. A step in the right direction for solving this is to create statistical methods which aim to identify and counter the root causes. I propose this can be done with more specific and complex causal and counterfactual models to infer multiple causes, structures and to avoid assumptions about social categories. I propose the possibility of extending this to meta-learning to further understand structures of inequality, which can also act as a technical basis for policy making and audits.Machine learning is highly effective in predicting outcomes accurately, thus providing the opportunity to allocate scarce societal resources quickly and efficiently. Consequently, machine learning has rapidly acquired a presence in high stakes decisions in socio-technical systems, which are systems that involve complex interactions between humans, machines and society. As machine learning has advanced in this space, its presence in the criminal justice system, health care, and the education system, shows that these algorithms were readily reproducing and exaggerating discrimination that exists in the world, causing significant harm.This led to the development of a new field in machine learning - fairness, with conferences such as Fairness, Accountability and Transparency (FAccT), and Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO) being created. Thus, in recent years, machine learning researchers have made significant effort to mitigate biases and discrimination exacerbated by its technologies, however there are significant gaps and failures within this current research area. Fairness algorithms are not generalisable beyond specific contexts and social inequalities are persistent, systemic and complex which is not reflected in the technical work. Little has been done to integrate the social sciences beyond the basic a priori argument of bias, and the majority of fairness work acts as a quick-fix in quality assurance, as opposed to trying to get to the root of the cause.This research proposal includes questions and ideas on how to integrate the complexity of social inequality, such as intersectional theory and infra-marginality, into statistics and machine learning, for a version of fair machine learning which correctly works with the complexes of society. While my research questions have been shaped by the wealth of previous work in this field, my specific questions are primarily based on work which aims for more complexity in causal models for fairness, such as, intersectionality in fair ranking and impact remediation. The main questions I want to address within my PhD work are:Can a more complex notion of fairness in machine learning be developed and understood with specific causal and counterfactual models to infer multiple causes and structures?Can this be combined with meta-learning to learn several algorithms, and thus learn several different discrimination hierarchies within the society?