Extending the explainability of machine learning models in policy decision making

政府和政策制定者越来越多地使用机器学习（ML）来支持决策。ML算法的性能通常随着模型复杂度的增加而提高，这使得最终用户更难查询模型输出。在模型的结构和支撑预测的关键变量方面缺乏透明度和问责制，可能导致不信任。人们对ML在关键决策中的迅速扩大使用表示担忧，特别是在影响边缘化部门和社区的政策方面。虽然近年来可解释的ML领域已经扩大，但现有的方法通常是“通用的”，无法捕捉真实世界最终用户的特定需求。因此，在不了解领域知识和特定需求/目标的情况下，现有可解释的ML方法的有效性仍然不清楚。了解可解释的ML的具体需求在政策决策中是非常必要的，因为政策是追求不同目标的人之间的妥协。这个博士项目旨在通过开发一个新的过程和框架来弥合差距，以确保ML模型可以更好地理解，从而更容易被政策制定者采用。我们重点关注两个方面。首先，我们通过考虑ML作为决策支持工具而不是预测工具的应用来解决这个问题。为了做到这一点，在开发ML模型时，我们明确地捕获决策者的观点，并确保它以他们可以理解，修改和询问的方式正式捕获在模型中。其次，我们使用通常用于解决棘手问题的可视化工具，例如因果图，为决策者捕获总体ML过程，以便更好地理解和解释整个过程。通过混合这两种方法，定量和定性，在开发可解释ML的框架方面将取得有意义的进展。

Governments and policy makers are increasing their use of machine learning (ML) to support decision-making. The performance of ML algorithms generally improves with the increase of model complexity, which makes it harder for end-users to interrogate the model output. Lack of transparency and accountability in how the model is structured and the key variables underpinning predictions can lead to mistrust. Concerns about the rapidly expanding use of ML in making critical decisions have been voiced, especially for policies affecting marginalised sectors and communities. While the area of explainable ML has expanded in recent years, the existing methods are often "general-purpose" that fail to capture the specific needs of real-world end-users. As such, the effectiveness of the existing explainable ML approaches remains unclear without understanding the domain knowledge and specific requirements/goals. Understanding the specific needs for explainable ML is highly demanded in policy decision-making since the policy emerges as a compromise between people pursuing different goals.This PhD project aims to bridge the gap by developing a novel process and framework to ensure that ML models can be better understood, and therefore more readily adopted by policy makers. We focus on two aspects. First, we approach the problem by considering the application of ML as a decision-support tool, rather than a predictive tool. To do this, when developing the ML models, we explicitly capture the views of the decision-maker and make sure it is formally captured in the models in a way they can understand, modify and interrogate. Second, we use visual tools commonly deployed for wicked problems such as causal maps to capture the overarching ML process for a decision-maker so that the overall process is better understood and explained. By mixing these two approaches, quantitative and qualitative, meaningful progress will be achieved in developing a framework on explainable ML.