Juraj Bodík

In an era of abundant and complex data, understanding "what causes what" is more crucial than ever. Despite massive datasets, causality cannot be established by black-box machine learning methods alone. My research develops formal mathematical frameworks that flexibly integrate domain knowledge into inference about unidentifiable causal quantities. I focus on three interconnected questions:

1. How can we reason causally beyond the observed range of the data? I study extrapolation to extreme or unseen treatment levels, such as a patient receiving a dosage far beyond recorded values, where the reliability of such reasoning depends critically on how extrapolation is modeled; a difficulty that becomes especially pronounced in causal settings.
2. How can we reason about non-identifiable quantities such as ITEs and counterfactuals? Although bounds for these quantities exist, they are often too wide to be informative in practice. To obtain sharper and more interpretable results, I introduced a new class of cross-world assumptions that make inference about such quantities feasible and practically useful.
3. In causal discovery, how can we relax restrictive yet necessary assumptions such as linearity when estimating the underlying causal graph? I develop frameworks that replace these with more interpretable and structurally grounded assumptions about how the data are generated.