School of Computer Science and Statistics
Eoin Delaney This project investigates methods for explaining anomalies detected in time series data. The aim is not only to identify when an anomaly occurs, but also to provide understandable explanations of why the observation or time period was considered anomalous. The project may explore feature importance, subsequence explanations, prototype examples, counterfactuals, or other explainable AI techniques … Read more
This project evaluates guardrail or safety models designed to detect, prevent, or filter unsafe and undesirable outputs from large language models (LLMs). The focus is on developing a systematic evaluation framework and comparing the effectiveness of different guardrail approaches. This project would be ideal for BA ICS/MCS, BA CS (JH), BA CSLL and integrated masters. … Read more
This project investigates algorithmic bias and fairness in machine learning systems. The student will analyse whether a predictive model produces unequal outcomes for different demographic or protected groups and evaluate techniques for measuring and mitigating unfairness. A focus will be on using and extending the OxonFair python toolkit. This project would be ideal for BA … Read more
Project descriptor: This project explores how machine learning models can estimate uncertainty in their predictions and use a reject option, allowing the model to abstain from making predictions when confidence is too low. The aim is to improve the reliability and safety of automated decision-making systems. This project would be ideal for BA ICS/MCS, BA … Read more
This project investigates interpretable machine learning methods that explain model predictions using examples. The focus will be on prototypes, which represent typical examples of a class or decision, and counterfactuals, which show how an input would need to change to receive a different prediction. Students will explore and compare state-of-the-art (SOTA) approaches. Interpretability in prediction … Read more
Risk detection models (such as IBM’s Granite Guardian) are increasingly used to flag harmful prompts and responses in large language model pipelines. These systems are trained on human and synthetic data to identify risks across multiple dimensions, but their reliability and fairness are not guaranteed. They may over-flag certain groups, miss subtle harms, or be … Read more
This project investigates how post-hoc counterfactual explanations can be used to debug opaque models such as deep neural networks by revealing which feature changes most influence predictions. In applications like anomaly detection, counterfactuals help clarify why certain cases are flagged as abnormal and expose when models rely on spurious correlations or biased patterns. By using … Read more
This project focuses on the rich field of algorithmic fairness where the goal is to ensure that predictions are not biased against subgroups of the population whilst maximising predictive performance. One challenge is when we focus on multiple protected attributes.