Developing Trust in Algorithmically Driven Services by Enhancing Explainable and Fair Machine Learning Through Co-Creation and Customer Interactions

目标和预期影响研究问题：在多大程度上可以使用与最终用户的共同创造方法来创新FairML/XAI？交互设计如何影响最终用户对FairML和/或XAI算法决策的信任。从最终用户和金融机构的角度来看，这些创新带来了哪些改进。工作成果和贡献：了解最新技术和待解决的问题在实验室和部署环境中评估以用户为中心的FairML/XAIFairML/XAI创新的方法评估和设计FairML实践很困难，原因有三：首先，要给公平下一个合适的定义是很困难的--伟大的思想家们已经为此争论了几个世纪--即使在专注的实用主义语境中，也经常存在复杂性和基于价值的紧张关系[1];其次，开发一种在整个社会中表现公平的数据驱动算法已被证明是例外而不是常态，近年来，有许多这样的例子可以出错[2]-[5];第三，这种自动化决策系统的部署、采用和对最终用户的影响往往是次要的考虑因素，或者是事后评估，如果有的话。博士年度计划大纲在第一年，工作将涉及博士研究人员沉浸在最先进的接口和与FairML和XAI算法等的交互方面。此外，他们将通过调查一系列现有的金融科技服务来进行初步研究，以记录和验证当前部署的方法。这项工作将进一步涉及焦点小组，研讨会，与客户和潜在客户的调查，以了解他们对偏见/缺乏公平性等的任何看法。第二年将开始进行一系列调查，试图找出改善FairML服务方法与客户之间建立关系的公平性和公正性的潜在方法：这项工作将试图梳理出人们在多大程度上相信服务机构公平、无偏见地对待他们。这可能涉及一系列经验技术，例如：用户与自动化服务交互的向导（实际上是由人类控制的），并主动询问他们是否理解和信任决策/互动（以及服务如何改进它）;向最终用户公开FairML模型的各种组件或服务决策的XAI解释的原型;第三年，将允许对第二年进行的实验工作进行有力的后续、重复和改进，进一步接触真实的世界背景。第一年和第二年将使博士研究人员能够收集一系列数据，然后可以用来考虑导致性能改进和增加偏见或公平感的关键因素的模式。如果成功，博士将能够展示一套工具和/或设计方法，组织可以用来发现这些细微差别的偏见，并使用它们来评估/改进他们的模型。第一年和第二年可以进行试点实验来测试这一假设，第三年用于更实质性的测试和论文的完整撰写。

Aims and Intended ImpactResearch Questions:To what extent can co-creation methods with end-users be used to innovate FairML/XAI?How does interaction design influence the trust end-user's have in FairML and/or XAI algorithm decision making.What improvements do such innovations bring from an end-user and Financial Institution point of view.Outcomes and contributions of the work:Understanding of state of the art and issues to be addressedMethods for user-centered FairML/XAIFairML/XAI innovations evaluated in lab and deployed settingsEvaluating and designing FairML practices is difficult for three broad reasons: First, arriving at a suitable definition of fairness is hard - great thinkers have debated about this for centuries - even in focused pragmatic contexts there are often complexities and value-based tensions [1]; Second, developing a data-driven algorithm that performs fairly across society has been shown to be the exception rather than the norm, recent years have been littered with many examples of how this can go wrong [2]-[5]; Third, the deployment, uptake and impact on the end-user of such automated decision systems is too often a secondary consideration or retrospectively appraised, if at all.PhD yearly plan outlineIn the first year, the work will involve the PhD researcher immersing themselves in the state-of-the art in terms of interfaces and interactions with FairML and XAI algorithms etc. Inaddition, they will carry out primary research by surveying a range of existing fintech services to document and characterise the approaches currently deployed. This work will furtherinvolve focus groups, workshops, surveys with customers and potential customers to understand any perceptions they have of bias/ lack of fairness etc.Year two will begin with a series of investigations that attempt to draw out potential ways of improving the fairness and justice of the relationship being built between FairML approachesto services and customers: this work would attempt to tease out the extent to which people believe services are treating them fairly and without bias. This might involve a range of empirical techniques such as: wizard-of-oz where users interact with an automated service (which is in fact being controlled by a human) and where they are proactively asked about whether they understand and trust decisions/ interactions (and how the service could improve it); prototypes that expose various components of FairML models or XAI explanations of service decisions to end-users; communities to supplement answers/ decisions to be contextualized; a-b style experiments where users responses are analysed and compared etc.Year three, will allow for robust follow-up, iteration and refinement of the experimental work carried out in year two with further exposure to the real world-context. Year 1 and year 2 will enable the PhD researcher to gather a range of data that could be then used to consider patterns that surface key factors that lead to performance improvements and increasedperception of bias or fairness. If successful, the PhD will be able to then demonstrate a suite of tools and/or design approaches that organizations can use to uncover these nuanced biases and use them to assess/ improve their models. Year 1 and 2 can carry out a pilot experiment to test this hypothesis with Year 3 being used for a more substantial test and full write up of the thesis.