Generation of Synthetic Continuous Numerical Data Using Generative Adversarial Networks

Abstract

Continuous numerical is a type of data which often used for unsupervised learning such as clustering. However, this valuable data often provided in a small amount because it is hard to obtain, expensive, required an expert to collect them, or not available because it contains confidential information that cannot be published. These limited data situations can be an obstacle for processing and analyzing data or restrain clustering related research in general. Therefore, there is a need to be an alternative that can replace or increase the amount of data. The proposed method is generating synthetic continuous numerical data using Generative Adversarial Networks (GANs). This study used two GAN architectures (GAN and CGAN) and focused on unlabeled continuous numerical data to provide replacement or additional data for the clustering task. The Quality of synthetic data was measured using the accuracy of the xgboost algorithm in classifying real and synthetic data. When the xgboost accuracy of perfectly realistic data is 50%, synthetic data based on CGAN achieving 63%. The result of this study shows that GAN can generate data similar enough and not significantly different from the real data.