A Unifying Perspective on Causal World Models: From Observations to Representations to Structure
TL;DR: Connects causal observations, representations, and structure in a unified account of the components needed for causal world models.
My research sits at the intersection of causal discovery, causal representation learning, computational argumentation, and trustworthy AI. I study how causal structure can be learned from data and knowledge, represented in ways that support reasoning, and made open to explanation, challenge, and revision. The broader goal is to develop causal AI that helps people make better decisions in high-stakes domains such as healthcare, finance, and public policy.
My PhD thesis, supervised by Prof. Francesca Toni and titled Causal Discovery for Trustworthy Artificial Intelligence, introduced methods that use causal discovery as a tool for explainability and contestability in AI. I combined game-theoretic concepts with computational argumentation to improve causal discovery and make its outputs more reliable, transparent, and open to human scrutiny and revision. The central contribution is Causal ABA, an Assumption-Based Argumentation framework that brings statistical evidence and domain knowledge into a common defeasible formalism, exposes conflicts between causal claims, and keeps people at the centre by allowing them to inspect, challenge, and revise the resulting causal models.
I subsequently served as co-investigator and technical lead on the ERC Proof of Concept project CArLA: Causal Argumentative Learning Assistant. The completed project developed Causal ABA into an interactive framework for transparent, explainable, and contestable causal discovery, combining attributable evidence from data, language models, and domain experts. Its outputs and publications are available on the project website, and the CArLA demonstration shows how users can inspect evidence, challenge causal claims, resolve conflicts, and trace the resulting revisions.
My current research focuses on three connected directions:
Together, these strands aim to connect statistical causal learning with symbolic reasoning, producing causal models whose evidence, mechanisms, and limitations remain visible to the people who rely on them.
TL;DR: Connects causal observations, representations, and structure in a unified account of the components needed for causal world models.
TL;DR: Studies when private causal reports identify a shared causal backbone and what additional information is needed when they do not.
TL;DR: Combines LLM-derived causal knowledge with statistical evidence through argumentation, keeping conflicts visible and inspectable.
TL;DR: Formalises how agents with private causal models can identify and communicate their shared causal backbone.
TL;DR: Uses optimal correction sets to resolve inconsistent causal constraints within argumentative causal discovery.
TL;DR: Models assumption-based argumentation as a heterogeneous graph and learns to predict whether assumptions are accepted.
TL;DR: Introduces gradual semantics for assumption-based argumentation, assigning degrees of acceptability rather than only binary outcomes.
TL;DR: Uses Shapley-inspired scores to make constraint-based causal discovery more robust to noisy conditional-independence evidence.
TL;DR: Explores heterogeneous graph neural networks for predicting credulous acceptance of assumptions in ABA frameworks.
TL;DR: Develops human-centred causal discovery methods that use game-theoretic ideas and argumentation to make AI explanations contestable and trustworthy.
TL;DR: Introduces an argumentation-based causal discovery framework that reconciles conflicting statistical evidence and domain knowledge.
TL;DR: Argues that computational argumentation provides the dialogue, justification, and revision mechanisms needed for genuinely contestable AI.
TL;DR: Surveys contemporary argumentation-for-AI methods, tools, and applications in an openly accessible handbook.
TL;DR: Uses causal discovery and knowledge injection to expose and revise neural-network mechanisms with expert input.
TL;DR: Presents an interactive causal discovery approach in which argumentation lets people inspect and challenge causal conclusions.
TL;DR: Builds argumentative explanations of classifier outputs from causal models, making supporting and opposing reasons explicit.
TL;DR: Transforms causal models into structured argumentative explanations that make the reasons for an outcome explicit.