Towards trustworthy AI: from local explanations to causal understanding

Thesis

In an era of rapidly advancing AI capabilities and mounting ethical scrutiny, the demand for transparent and trustworthy machine learning solutions has never been more urgent. High-stakes domains such as healthcare, criminal justice, and financial services increasingly deploy complex AI systems that can profoundly impact human lives, yet their opaque nature often renders their decisions inscrutable to the very individuals affected by them. This lack of transparency not only raises ethical concerns but also poses practical challenges for practitioners who must ensure these systems behave as intended when deployed in real-world settings.

This dissertation explores the challenge of approximating different types of truth in machine learning, offering methodological improvements that advance both local explainability and causal inference. While XAI methods seek to illuminate the inner workings of black-box models, causal inference techniques aim to uncover the underlying mechanisms that drive observed phenomena. Together, these approaches form a comprehensive framework for understanding - XAI explains our models of reality, while causal inference explains reality itself. Throughout this work, I focus on improving approximations of various truth concepts: causal truths, model truths, and true data distributions.

My research begins by addressing foundational limitations in local model explanations, introducing Neighbourhood SHAP values that leverage local reference distributions to provide more meaningful feature attributions that better reflect local model behavior while demonstrating increased robustness against adversarial classifiers. Building on this foundation, I develop Path-Wise Shapley effects (PWSHAP), bridging predictive modeling and causal understanding by combining user-defined causal structures with Shapley values to assess variable impacts through specific causal pathways. Moving beyond model explanations to direct causal inference, I introduce BICauseTree, an interpretable balancing method that identifies clusters where natural experiments occur locally, detecting sub-groups with positivity violations while providing explicit definitions of valid inference populations. Finally, I address the challenge of securely evaluating dataset value for causal investigations, developing an Expected Information Gain framework that allows data hosts to assess merge potential without compromising sensitive information. Collectively, these methodological innovations advance both theoretical understanding and practical applications in explainable and causal AI, offering solutions that enhance transparency, improve decision support, and strengthen trust in AI systems deployed in high-stakes domains.

Authors: Ter-Minassian, L

• DOI: 10.5287/ora-er452zx86
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: causal inference; explainable AI; robust machine learning; Bayesian statistics
• Subjects: Statistics; Machine learning