Improving Gradient-based DAG Learning by Structural Asymmetry

Directed acyclic graph (DAG) learning plays a fun-damental role in causal inference and other scientific scenes, which aims to uncover the relationships between variables. However, identifying a DAG from observational data has al-ways been a challenging task. Recently, gradient-based DAG learning algorithms that convert a combination-optimization DAG learning problem into a continuous-optimization problem have achieved emerging successes. These algorithms are easy to optimize and able to deal with both parametric and non-parametric data but suffer from many reversed edges learnt by these algorithms. In this paper, we propose a framework named Residual Independence Test (RIT) to correct those reversed edges by leveraging the structural asymmetry reflected in the depen-dence between regression residual and direct cause. We conduct extensive experiments on both synthetic and benchmark datasets, the results show that the RIT framework significantly improve the performance of gradient-based DAG learning algorithms.