Causal Inference & Transportability
Robust causal abstractions and transferring causal conclusions across populations.
Causal models express how a system responds to interventions. My research develops abstractions that relate these models across levels of detail and studies when causal conclusions remain useful across changing environments and populations.
Distributionally robust causal abstractions
Our ICML 2026 work develops causal abstraction methods that account for environmental shifts and model misspecification. We formulate learning as a constrained min-max problem with Wasserstein ambiguity sets and derive guarantees on worst-case abstraction error.
This provides a way to reason about the reliability of an abstraction when the assumed environment or model is imperfect.
Generalised transportability
Our transportability work takes a model-level perspective: can one abstraction map align the interventional behaviour of a source and target population?
When exact alignment is unavailable, we derive certified intervals for causal queries. The framework also addresses settings where a query is not exactly transportable or target data are unavailable, under explicit assumptions on mechanism and environment perturbations.
Selected publications
- Yorgos Felekis, Theodoros Damoulas, and Paris Giampouras. Distributionally Robust Causal Abstractions. ICML 2026.
- Yorgos Felekis, Paris Giampouras, Fabio Massimo Zennaro, and Theodoros Damoulas. Generalised Transportability via Causal Abstractions. arXiv, 2026.