Multi_TimeGAN

Multivariate Time Series Imputation with Generative Adversarial Networks

This is original repository is :

https://github.com/Luoyonghong/Multivariate-Time-Series-Imputation-with-Generative-Adversarial-Networks

Special thank to Yonghong Luo for his great ideas and contribution.

Fine-tuning version for learning purpose

author: Yonghong Luo, Xiangrui Cai, Ying Zhang, Jun Xu and Xiaojie Yuan

tensorflow version:1.7 python:2.7

The proposed method is a two-stage method. We first train GAN, then we train the input vector of the generator of GAN.

To run the code, go to the Gan_Imputation folder:

Execute the Physionet_main.py file, then we will get 3 folders named as "checkpoint" (the saved models), G_results (the generated samples), imputation_test_results (the imputed test dataset) and imputation_train_results (the imputed train dataset).

Go to GRUI floder

Excute the Run_GAN_imputed.py file, then one floder-"checkpoint_physionet_imputed" will be created, go to the "checkpoint_physionet_imputed/30_8_128_64_0.001_400_True_True_True_0.15_0.5" floder, find "result" file, the "result" file stands for the mortality prediction results by The RNN classifier trained on the GAN imputed dataset. The first column is epoch, the second column is accuracy and the last column is the AUC score.

Final result file location

GRUI/max_auc is the file that record final auc score