New Method Learns Causal Structures in Clustered, Heterogeneous Data

Recent advancements have made learning Directed Acyclic Graph (DAG) structures, crucial for causal discovery, scalable. However, existing techniques typically assume a homogeneous population, making them unsuitable for clustered data where variations specific to individual clusters, such as patient-specific effects, are common. This research addresses this limitation by proposing a novel approach for structure learning in such heterogeneous environments. The core idea is to extend the classical fixed- and random-effects framework, commonly used in mixed models, to the domain of structure learning. This allows the method to estimate a global causal structure while simultaneously accounting for unique local effects within each cluster. A key technical contribution is a differentiable graph coupling mechanism that mathematically guarantees the combined fixed- and random-effects graphs remain acyclic, a fundamental requirement for causal inference. Computationally, the researchers provide a provably convergent first-order method that leverages efficient batched updates across clusters. Statistically, the model's identifiability is established, and it is shown to asymptotically recover the true structure. Experiments on both real and synthetic data demonstrate that this new approach can detect dependencies that other estimators miss, underscoring its significant value for causal structure learning in clustered settings.