Roman Marchant

Associate Professor Roman Marchant is Head of Research for the Thrive Program at the Human Technology Institute, University of Technology Sydney. He specialises in probabilistic machine learning and Bayesian methods for causal inference and decision-making under uncertainty. Roman has led interdisciplinary teams across education, health, and public policy, developing ethical, data-driven approaches to complex social challenges. He holds a PhD in computer science and has published on decision-making under uncertainty in fields including robotics, environmental monitoring, criminology, mental and physical health, and education—most recently proposing a sequential methodology for Bayesian adaptive trials in public policy. Roman has taught postgraduate courses on Probabilistic ML, Bayesian Inference, and AI Ethics, convened Australia’s first Ethics of Data Science Conference, and serves as Associate Editor for the journal Data & Policy.

Talk Title

Bayesian Structure Learning: Empowering Policy Through Causal Inference

Talk Summary

This talk explores how Bayesian Structure Learning supports evidence-based policy, drawing on research from the Human Technology Institute (HTI) and the Thrive program. I begin by introducing our methodological cycle, which integrates Bayesian Networks (BNs), structure learning, expert knowledge, and community co-design to inform decisions in complex social settings.

Assuming familiarity with BNs, I focus on recent advances in causal discovery—covering models, inference algorithms, and sequential decision-making strategies. Case studies in childhood obesity, mental health, and education demonstrate how these methods reveal actionable distinctions between proximal and upstream causes, enabling more targeted interventions.

I will also discuss key methodological challenges, such as population heterogeneity and the need for mixtures of BNs to yield nuanced, individualised insights. In addition, I highlight the emerging role of Large Language Models (LLMs) in leveraging unstructured text data to inform causal models. I conclude with a forward-looking view on the role of Bayesian approaches in creating adaptive, context-sensitive policy tools that are both rigorous and socially responsive.