New framework integrates expert knowledge to boost causal discovery speed and accuracy

Causal discovery—inferring cause-effect relationships from data—is computationally expensive for large numbers of variables. Expert background knowledge (e.g., known constraints or prior findings) is often available but typically only applied after discovery to refine the resulting graph. This post-hoc approach misses opportunities to shrink the search space early, wasting compute and potentially missing better structures. Schubert and colleagues introduce a framework that integrates such knowledge directly into the discovery process, specifically targeting scalable methods that recover only a subset of the full graph (e.g., local or subgraph discovery).

Their framework is agnostic to the underlying algorithm, and they demonstrate it on several popular causal discovery methods. Empirical results show that using background knowledge during search reduces the number of candidate graphs explored, leading to faster runtimes and higher structural accuracy compared to standard post-processing. The work addresses a key bottleneck in applying causal discovery to real-world domains like genomics, economics, and social science, where expert constraints are abundant but underutilized. The paper is available on arXiv (2607.10456).