Peering Inside the Black Box: Exploring Innovations in Interpretable Machine Learning and Causal Inference for the Explanation of Political Violence

机器学习（ML）在社会科学研究中的应用越来越普遍。然而，在我们如何在推理环境中使用这些模型来解释而不是预测或分类现象方面仍然存在差距。与此同时，决策越来越需要理解越来越常见但复杂的机器学习策略。尽管ML模型能够比传统统计模型更大程度地捕捉社会现象的复杂性，但它们被认为对解释和推理不太有用，部分原因是被视为“黑箱”方法，基本上是不透明的。最近的方法学进展表明，我们可以理解这些复杂的ML模型，并将它们用于解释和因果推理。这些进展被归类为“可解释机器学习”（IML）和“因果机器学习”（Causal ML）的子领域。到目前为止，这些创新还没有在社会科学中得到充分的转化。这项研究旨在填补这一空白，并展示使用ML方法不仅用于预测，而且用于解释和推理的实用性。我将通过翻译、创新以及最终将这些方法应用于政治暴力的研究来做到这一点。研究ML方法在政治暴力中的应用尤为重要，因为了解冲突的决定因素可以采取各种形式的积极干预和政策解决方案。然而，政治暴力是复杂的，多原因的，因此传统的参数方法是不可能有效地模拟这些动态。就决策相关性而言，我们必须能够解释为什么会发生冲突事件，以及是否以及何时发生。在方法论上，我的目标是展示研究人员如何使用ML来捕捉和解释复杂的社会现象，并在此过程中为更大的决策相关性和影响力开辟途径。那么，一个可能适用于这个博士项目的广泛研究问题是“我们如何利用复杂机器学习策略的发展来进行社会科学背景下的解释和推理？“实质上，我的目标是利用这些工具来促进我们对政治暴力机制的理解。

The use of machine learning (ML) within social science research is increasingly common. Yet, a gap remains in how we use these models in inference settings to explain, rather than predict or classify, phenomena. Concurrently, there is a growing need in policymaking to make sense of increasingly common, but complex, machine learning strategies. Despite ML models being able to capture the complexity of social phenomena to a greater extent than traditional statistical models, they have been considered less useful for explanations and inferences, in part due to being viewed as 'black box' methods that are substantively opaque. Recent methodological advances show that we can make sense of these complex ML models and use them for both explanation and causal inferences. These advances are grouped under the subfields of 'Interpretable machine learning' (IML) and 'Causal machine learning' (Causal ML). To date, these innovations have not been sufficiently translated within the social sciences. This research aims to fill this gap and demonstrate the utility of using ML methods not just for prediction, but for explanation and inference also. I will do so by engaging in translation, innovation, and finally application of these methods to the study of political violence. Studying the application of ML methods in political violence is particularly important as understanding the determinants of conflict allow proactive interventions and policy solutions in various forms. Political violence is, however, complex, and multi-causal, and therefore conventional parametric methods are unlikely to model these dynamics effectively. For policymaking relevance, it is imperative that we can explain why a conflict event might occur as well as if and when. Methodologically, I aim to demonstrate how researchers can use ML to capture and explain complex social phenomena, and in the process open avenues for greater policymaking relevance and impact. A broad research question that could apply to this PhD project, then, is "How can we leverage developments in complex machine learning strategies for explanation and inference in social science contexts?" Substantively, I aim to use these tools to advance our understanding of the mechanisms of political violence.