Generative Adversarial Nets For Fragmentary Data Im-Putation And Prediction

Fragmentary data is often encountered in modern scientific research and application.It has a high data loss rate and complex response mode,which brings great challenges to the task of imputation and label prediction.Existing statistical methods can provide useful theoretical properties when dealing with fragmentary data,but they usually need to rely on some model assumptions,and the types of data that can be processed are not flexible enough.On the other hand,the method based on generative adversarial nets in the field of machine learning either has no theoretical guarantee,or only considers the case of missing completely at random mechanism.And most of them separate the two tasks of imputation and label prediction,which affects the effect of label prediction task.By using the structural characteristics of fragmentary data response mode,this paper proposes a new method based on generative adversarial nets-fragmgGAN to impute and predict the fragmentary data.It can flexibly handle various data types.Under the assumption of missing at random,we prove that the imputed data distribution is consistent with the real data distribution,and this conclusion does not depend on any model assumption of the data.At the same time,this paper also extends the theoretical results of previous work.FragmGAN uses both generator and discriminator to train predictor.This linkage mechanism shows significant advantages in prediction performance in a large number of experiments.