On Causal Explainability in Machine Learning: Philosophy, Methods and Applications

The project aims to improve and develop the explanation methods used explaining the outputs and predictions of machine learning models through interdisciplinary and complimentary methods in philosophy and data science. A particular motivation of improving explainability for the project is about being able to adequately explain the fairness of a model. The project aims to answer how explanations should be designed and used to improve understanding by using topics in the philosophy of science such as idealization and causal explanation. The primary philosophical theory being developed in the project is concerned with how each explanation method forms a distinct idealization, that these idealizations are often improved the more causal information they have because this allows for more robust intervention, and that the idealization framework helps to form a taxonomy of explanation methods that would define a machine learning task. The project further aims to evaluate and improve existing explanation methods by employing causal reasoning and causal methods, for example by creating causal data simulations to compare ground truths to explanations as a form of evaluating existing popular explanation methods such as partial dependence plots (PDP) and Shapley explanations (SHAP), and by adding causal knowledge to popular XAI methods to create new methods, such as causal dependence plots inspired by partial dependence plots. The project aims in this sense to promote explanations for machine learning algorithms that can do more than just measure direct effects between features, but also other causal effects that are of interest such as total dependence. To assess the efficacy of causal explanation in applications, the project seeks to access data on UK crime sentencing (Data First Crown Court and Magistrates Linking Dataset) and apply machine learning and causal explainability to analyse ethnic disparities in sentencing.