Robust counterfactual explanations in machine learning
File(s)
Author(s)
Jiang, Junqi
Type
Thesis or dissertation
Abstract
Recent advances in machine learning (ML) have led to the widespread deployment of complex black-box models, raising concerns about their trustworthiness. Explainable artificial intelligence (XAI) aims to address these concerns. Within XAI, counterfactual explanations (CEs) have emerged as a prominent paradigm, providing insights into how an input must be modified to alter a model's prediction. CEs are highly valuable in high-stakes domains such as finance and healthcare, where they enable algorithmic recourse, informing users negatively affected by model decisions how to achieve more favourable outcomes. Beyond fundamental requirements like validity and proximity, CEs must ideally be plausible and diverse. However, current CE methods often lack robustness: small perturbations in the generation process can render a previously valid CE invalid, posing significant deployment risks.
This thesis investigates the robustness of CEs. We define novel forms of robustness theoretically, characterise scenarios where robustness is lacking, and introduce methods improving upon the state-of-the-art. Specifically, we study three forms of robustness: against model changes, input changes, and model multiplicity. First, we propose a technique to certify CEs' robustness against changes in the ML model. We then propose new methods that jointly encourage provable robustness and plausibility, advancing state-of-the-art performance across various CE metrics. Next, we address robustness against input changes, ensuring similar inputs receive similar CEs. We propose a comprehensive model-agnostic method to achieve this, leveraging a novel generative model that learns a structured latent space where each class label is represented by Gaussian components with diverse, prototypical centroids. Finally, we study CE robustness under model multiplicity, where multiple trained models may produce conflicting predictions. We formalise this setting as an ensembling problem and develop a method providing robust recourse. Collectively, these contributions advance the theoretical and practical aspects of CE robustness, offering new approaches for building more trustworthy XAI systems.
This thesis investigates the robustness of CEs. We define novel forms of robustness theoretically, characterise scenarios where robustness is lacking, and introduce methods improving upon the state-of-the-art. Specifically, we study three forms of robustness: against model changes, input changes, and model multiplicity. First, we propose a technique to certify CEs' robustness against changes in the ML model. We then propose new methods that jointly encourage provable robustness and plausibility, advancing state-of-the-art performance across various CE metrics. Next, we address robustness against input changes, ensuring similar inputs receive similar CEs. We propose a comprehensive model-agnostic method to achieve this, leveraging a novel generative model that learns a structured latent space where each class label is represented by Gaussian components with diverse, prototypical centroids. Finally, we study CE robustness under model multiplicity, where multiple trained models may produce conflicting predictions. We formalise this setting as an ensembling problem and develop a method providing robust recourse. Collectively, these contributions advance the theoretical and practical aspects of CE robustness, offering new approaches for building more trustworthy XAI systems.
Version
Open Access
Date Issued
2025-10-03
Date Awarded
2026-09-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Toni, Francesca
Leofante, Francesco
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)