Causal Structure Leraning Based On Generative Adversarial Networks And Constraint Algorithms

Studying causality is the ladder of scientific progress,and mining causality in data is universal to all disciplines.We are in the era of big data,finding causal relationships based on observational data has become the driving force behind the development of various fields.In recent years,more and more researchers have devoted themselves to finding ways to infer causality from passively observable data,but inferring structural graphs representing causality from observational data is extremely challenging and complex.Focusing on this theme,based on the idea of deep learning,this paper designs a novel causal structure learning method SNCIT-PC algorithm for inferring directed acyclic graphs representing causal relationships from observational data.In this paper,we first designs a conditional independent test method SNGCIT algorithm using spectral normalization generative adversarial networks.It is non-parametric,does not require assumptions about the data distribution,and can effectively perform conditional independent testing even when the confounding variables are high-dimensional.This paper also gives the asymptotic theoretical properties of the SNGCIT method,which guarantees the feasibility of the method theoretically.In addition,from the simulated data,this method can asymptotically control the probability of making Type I error while maintaining a high test power.Next,this paper corrects the order of edge deletion in the original PC algorithm and uses the SNGCIT method for conditional independence test.Compared with the original PC algorithm,the new method SNCIT-PC algorithm is no longer limited to only infer multivariate normal distribution data sets,but can be applied to more distribution types of data sets.In addition,in order to speed up the SNCIT-PC method and enable it to infer higher-dimensional datasets,this paper designs a scheme for multiple cores of the CPU to simultaneously detect conditional independence in the skeleton graph learning stage.Compared with the original SNCIT-PC algorithm,the calculation speed is increased by at least 4 times.Finally,this paper validates the superiority of the SNCIT-PC algorithm on simulated data and real protein signaling network data.From the numerical experimental results,as the sample size increases,the performance of the SNCIT-PC method improves? in addition,as the dimension increases,the SNCIT-PC algorithm can maintain a high F score.We also used the SNCIT-PC algorithm for the construction of protein signaling networks to identify potential causal relationships between proteins,providing insights for an accurate understanding of normal cellular responses as well as potential dysregulation of disease.This demonstrates the prospect that the SNCIT-PC method can be used to identify complex causal relationships between various molecules in biology.