Multivariate Time-Series Data Generation in Generative Adversarial Networks

Time-series data often arises during the monitoring and evaluation of ongoing industrial processes. Time series forecasting requires accurate data modelling through the description of inherent structures such as trend, cycle, and seasonality by collecting and modeling stochastically the historical data points of a time series. In this paper, we are concerned with industrial time series data that is limited and not readily available for accurate machine learning tasks, e.g., online fraud and network intrusion data. In this scenario, modeling of time series can be achieved through generative modeling activities in deep learning. Then, abundant temporal data can be generated and used in different ways to achieve application-level forecasts and predictions. We focus on the use of Generative Adversarial Networks (GANs) to model and generate limited real-world time-series data. We discover that this is a relatively new research domain with research trends generally focusing on employing real data to generate or forecast the time series through the GAN in a supervised manner. On the contrary, we adopt a novel approach that is completely unsupervised, i.e., we employ GAN to generate limited time series data from a (gaussian) noise distribution as input without any additional input vector of real data. To achieve realistic generative performance in this situation, we propose and implement a feedback mechanism through which GAN improves its performance by using historically generated time series (and never the real data). Using different experimental configurations, we demonstrate that our approach generates realistic limited intrusion detection data from the standard CIC-IDS2017 dataset.