CTGAN VS TGAN? Which one is more suitable for generating synthetic EEG data

Min Jong Cheon, Dong Hee Lee, Ji Woong Park, Hye Jin Choi, Jun Seuck Lee, Ook Lee

Research output: Contribution to journalArticlepeer-review


BCI has been an alternative method of communication between a user and a system, and EEG is a representative non-invasive neuroimaging technique in BCI research. However, gathering a large dataset of EEG is difficult due to insufficient conditions. Therefore, a data augmentation is required for the data and a generative adversarial network is a representative model for the augmentation. As the EEG data is a CSV format, we decided to utilize CTGAN and TGAN for creating synthetic data. Our research was conducted through 3 steps. First of all, we compared two datasets from each model through data visualization. Secondly, we conducted a statical method for calculating similarity score. Lastly, we used both data as input data of the machine learning algorithms. Through the first step and second step, we found that the data from CTGAN has higher similarity than TGAN. However, in the last step, the result showed that the result such as accuracy, precision, recall, f1 score showed no significant difference between the two datasets. Furthermore, compared to the original dataset, none of the synthetic datasets showed higher scores. Therefore, we concluded that further research is needed to find out a better method for data augmentation so that the synthetic data could be utilized for the input data of machine learning or deep learning algorithms.

Original languageEnglish
Pages (from-to)2359-2372
Number of pages14
JournalJournal of Theoretical and Applied Information Technology
Issue number10
StatePublished - 2021 May 31


  • Artificial Intelligence
  • BCI
  • Data Augmentation
  • Deep Learning
  • EEG
  • GAN


Dive into the research topics of 'CTGAN VS TGAN? Which one is more suitable for generating synthetic EEG data'. Together they form a unique fingerprint.

Cite this