Synthesizing Tabular Time Series Data using Transformers

Abstract

Using synthetic data in place of real data can come with numerous benefits, such as the protection of privacy. However, synthesizing tabular data is difficult, since it is heterogeneous and might contain relationships between its columns and between its rows. While there has been much work dedicated towards generating synthetic tabular data with independent rows based on real data, less has been done towards generating time series tabular data, as it contains an extra temporal component. One such work uses a transformer model, and we use this work as a baseline for our own work. We specifically created a service in order to address the problem of deploying the transformer model on the cloud and to increase accessibility to the transformer model, and looked into addressing the limitations of how the previous work utilized the model. In addition, we performed a case study on the architecture of our service, where we investigated a limitation of our architecture, explored metrics for evaluating the synthetic data output from the architecture, and looked into using the architecture towards generating a forecast, an application the original transformer model is not originally designed for.