This study addresses the challenge of recovering causal structures from contaminated data using anchored linear structural equation models (SEMs). It discusses two novel approaches to relax existing restrictive identifiability conditions. First, for a...
This study addresses the challenge of recovering causal structures from contaminated data using anchored linear structural equation models (SEMs). It discusses two novel approaches to relax existing restrictive identifiability conditions. First, for anchored Gaussian DAG models, the anchored-frugality assumption posits that the true graph is the most frugal among candidate structures satisfying the Markov condition. This allows for the identification of the Markov equivalence class without requiring prior knowledge of measurement error variances and leads to the development of the Frugal-PC algorithm. Second, for distribution-free settings, the geometry-faithfulness assumption ensures that partial correlations serve as direct indicators of d-separation regardless of error distributions. This establishes the identifiability of distribution-free anchored linear SEMs and validates the use of the standard PC algorithm with Fisher's z-test. Extensive experiments on synthetic and real datasets, including breast cancer and galaxy data, validate the effectiveness of these theoretical contributions.