FoundCause: Causal Discovery with Latent Confounders from Observational Data

FoundCause is a novel amortized causal discovery model designed to infer causal graphs from observational data, specifically addressing latent confounders without requiring interventions. It employs a permutation-invariant transformer architecture with features like statistics-conditioned attention and a triangular refinement module, allowing it to outperform 11 classical and 4 amortized methods on 15 real-world datasets, achieving a +9.6% improvement in F1 score and a significant reduction in structural Hamming distance. This advancement is crucial for practitioners as it simplifies causal inference in complex datasets and enhances the robustness of causal models in the presence of hidden confounders.