Automating Fair Decision-Making

Automated decision-making systems, which utilise machine learning models, increasingly shape access to credit, employment, healthcare, and public services, yet their predictions can inherit and amplify historical inequities. Fairness is hard to ensure in practice because protected attributes can influence decisions through proxy features, fairness definitions are often incompatible, and reducing disparity can degrade accuracy or other criteria. This thesis treats fairness as a pipeline-level problem and develops complementary mechanisms that act before, during, and after model training.

At the pre-processing stage, demographic cues in textual documents can be discovered by a model during training and subsequently used to bias decisions. This thesis proposes the use of large language models to rewrite user-generated text so that task-relevant meaning is preserved while demographic cues are reduced. This prompt-based transformation pipeline mitigates demographic leakage without modifying the downstream learner.

During training, a model may exploit protected attributes and their proxies using correlations between these features and outcomes, amplifying biases already present in the data. Therefore, at the in-processing stage, we propose a powerful debiasing objective that directly shapes the learned representation to reduce sensitive-attribute dependence while preserving task-relevant signal. The proposed model provides explicit control over the accuracy-fairness trade-off through a small set of complementary regularisation terms, enabling consistent improvements across multiple fairness criteria without committing the optimiser to a single parity metric.

Even when upstream mitigations are applied and group-level fairness metrics are satisfied, unfair outcomes can still occur at the individual level. We therefore treat such unfairness as a safety-level failure of the decision-making system. Accordingly, for post-processing we introduce an inference-time fairness monitor that reframes bias sensitivity as a deployment-time risk. This chapter combines evidential uncertainty modelling with counterfactual masking to quantify how much predictive belief relies on protected attributes, enabling selective flagging under coverage control and guiding downstream post-processing.