Mátyás Schubert

Causal discovery methods enable the identification of adjustment sets for estimating causal effects even when the underlying causal graph is unknown. However, in large-scale settings they often become computationally infeasible or discover statistically inefficient adjustment sets, limiting their practical applicability. This PhD project investigates how to make causal discovery more scalable while preserving statistical efficiency, focusing on settings where we only wish to estimate the causal effects between a small subset of target variables. To this end, it develops approaches that restrict discovery only to relevant substructures to recover statistically efficient adjustment sets comparable to those obtained from full-graph methods. The project further studies how these approaches can incorporate different forms of background knowledge into the discovery process to improve both accuracy and efficiency, and how they can be extended to allow for latent confounding variables. Ultimately, this work seeks to develop principled methods for sound and complete identification of causal relations and optimal or near-optimal adjustment sets that combine scalability, statistical optimality, and practical usability for complex real-world data.