A0537
Title: Learning a directed acyclic graph with additive heteroskedastic errors
Authors: Chunlin Li - University of Virginia and Iowa State University (United States) [presenting]
Abstract: A causal discovery method is introduced to learn the causal relations in a directed acyclic graph (DAG) with heteroscedastic errors. First, the model identifiability is derived for DAGs with heteroskedastic errors. Then, a new DAG reconstruction method is proposed called the residual quantile (ResQ) estimation, which iteratively reconstructs the causal order of the variables. It is proven that ResQ enjoys desirable statistical properties such as reconstruction consistency in low- and high-dimensional cases and estimation robustness to heteroskedastic/heavy-tailed/contaminated errors. The theoretical properties have been substantiated by the synthetic experiments and applications to real-world causal benchmark datasets, where ResQ compares favorably against state-of-the-art competitors.
