Applications Of The GANs In The Causal Inference

In recent years,the research on causal inference has been developing continuously,and has been fully practiced in the fields of natural science and social science.However,the traditional causal inference models have many shortcomings in practical applications.For example,the propensity score matching(PSM)model is based on a large sample.In the case of a small sample size,partial covariate imbalance still exists through PSM.Therefore,the PSM model cannot provide effective causal inference.At present,some scholars in the field of machine learning have proposed some new causal inference methods combined with Generative Adversarial Networks(GAN)to optimize the performance of the PSM model.On the basis of PSM,this paper proposes a new causal inference model: GPSM model.The core of the model lies in: GAN learns the distribution of the original dataset,generates more samples with a similar distribution,and combines the original samples and the generated samples for propensity score matching,so as to solve the limitations of the traditional PSM model in the case of small samples,and control the selectivity error.The GPSM model is mainly composed of two sub-models:The first sub-model is a sample generation model based on GAN.It is composed of two neural networks: generator G and discriminator D.Through the adversarial learning between generator and discriminator,new simulation samples are generated,so the number of samples is expanded.The output of the submodel is combined with the real samples as the input of the second submodel.The second sub-model is a sample matching model based on PSM.The model calculate the propensity scores of the samples by Logistic regression,and use the propensity scores to match the samples from treatment group and control group with the same characteristics.So the counterfactual state of the object from the treatment group is approximately the matched object from the control group.Then evaluate the causal effects for treated by comparing the result of two objects.Finally,the GPSM model was empirically analyzed against the experimental background of the effectiveness of a nephrotic drug.The GPSM model proposed in this paper was used to evaluate the effect of experimental drugs on the treatment of nephropathy,and the causal inference process and results of GPSM model and traditional PSM model were compared.According to the estimator of the treatment effect for the treated and the result of hypothesis test,the experimental drug has significant effect on treatment.The empirical results show that the GPSM model is better than the traditional PSM model in terms of covariate balancing and model robustness,and the comprehensive performance of the GPSM model is better with the increase of the size of generated samples.