In machine learning, models typically capture correlations between variables, allowing them to predict the value of one variable based on others if the data distribution stays the same. Causality takes it further by studying cause-and-effect relationships: how taking a drug influences the disease outcome, for example. Causal graphs are one way to represent these relationships between variables: variables are nodes, and direct causal relationships are the (directed) edges of the causal graphs. Causal graphs allow for the estimation of intervention effects, such as estimating whether taking a certain drug might reduce blood pressure.
Causal structure learning refers to learning these causal graphs from the data. Classical approaches (e.g., PC or LiNGAM) use statistical tests or optimization to infer which variables have a direct effect on others, while newer methods rely on deep learning, expanding their application area to high-dimensional data (e.g., many variables from health records or biological data).
The objective of this master’s project would be to review the existing causal structure learning methods, develop a framework to compare different structure learning methods across different simulated datasets where the true causal graphs are known, and compare a subset of reviewed algorithms across multiple simulated datasets.
Related literature:
Heinze-Deml, C., Maathuis, M. H., & Meinshausen, N. (2018). Causal Structure Learning. Annual Review of Statistics and Its Application, 5(1), 371–391 https://doi.org/10.1146/annurev-statistics-031017-100630
Textor, J. (2025). When will causal structure learning become practical? International Journal of Epidemiology, 54(5), dyaf145. https://doi.org/10.1093/ije/dyaf145