Generating synthetic time series data in case of data scarcity
| Title | Generating synthetic time series data in case of data scarcity |
|---|---|
| Summary | Generating synthetic time series data in case of data scarcity |
| Keywords | |
| TimeFrame | |
| References | - Using generative adversarial networks (GAN) to simulate central-place foraging trajectories
https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210X.13853 - Data-Driven Crowd Simulation with Generative Adversarial Networks http://rainbow-doc.irisa.fr/pdf/2019_amirian_casa.pdf |
| Prerequisites | |
| Author | |
| Supervisor | Alexander Galozy, Peyman Mashhadi |
| Level | Master |
| Status | Open |
This project addresses generating synthetic sequential data in case of real data scarcity. The idea is that there exists statistics about the real data but only very few data points. The goal of the project is to develop a generative method (for example GANs, transformers, variational autoencoder) to generate synthetic data with as similar as possible statistics to the real data. Then, the small sample from real data will be utilized to generate the trajectory of sequential data.
In many of the real-world industry data problems, we face we are only provided with a small number of data samples of time-series data, but data statistics are readily available. It is of high interest to diversify the little data we have to train better machine learning models. Further, we would like to generate realistic synthetic data beyond mere data augmentation, that is task specific, thus resulting in better sample complexity of our developed algorithms.
