Causal discovery is an important part of causal inference and the goal is to discover the data genera-tion mechanism in the form of Directed Acyclic Graphs(DAGs).With regard to causal discovery,existing methods rarely take into account the presence of missing values in observational data.However,incomplete datasets are ubiquitous in practical scenarios,and figuring out the causal relationships in incomplete datasets has become a critical issue to be solved.In this paper,a new Causal Feedback-based Imputation Causal Dis-covery(CF-ICD)algorithm is proposed to achieve causal discovery of incomplete data sets.Generative Ad-versarial Networks(GAN)are used to estimate the distribution of missing data.The causal learning module based on Actor-Critic is used to search the optimal DAG,and a custom reward function based on the ex-tended Bayesian Information Criteria(eBIC)is designed.Classification error is introduced to guide the model to accelerate the exploration process and improve the stability.Extensive experimental results on simu-lated data and real data show that the proposed method is superior to existing methods under different data missing rates.
Deep learningData imputationCausal discoveryDirected Acyclic Graph
国家科技期刊平台
NSTL
万方数据
