A1268
Title: Causal discovery for linear acyclic models with gaussian noise using ancestral relationships
Authors: Ming Cai - Graduate School of Informatics, Kyoto University (China) [presenting]
Penggang Gao - Kyoto University (China)
Hao Wang - Kyoto University (China)
Hisayuki Hara - Kyoto University (Japan)
Abstract: The PC algorithm proposed relies solely on the faithfulness assumption of the causal model and identifies causal structures up to the Markov equivalence class. In contrast, LiNGAM added the assumption of the linearity and continuous non-Gaussianity of disturbances to the causal model, which allows for the complete identification of the causal DAG. By integrating the strengths of both the PC algorithm and LiNGAM, PC-LiNGAM makes it possible to identify the causal structure up to the distributed equivalence pattern beyond the Markov equivalence class, even in the presence of some Gaussian disturbances. However, PC-LiNGAM suffers from high computational complexity, which is factorial about the number of variables in the worst case. A new algorithm that significantly reduces the time complexity for learning distributed equivalence patterns is introduced. The main idea of the proposed method is to use the causal ancestor finding of Maeda and Shimizu, which is generalized to include Gaussian disturbances.
