Advanced Methods for Time Series Data Augmentation

Modern machine learning methods are data-intensive processes, requiring massive amounts of training data to achieve a high level of performance. As such, these techniques are challenging to deploy on tasks where datasets are especially difficult or expensive to obtain or where edge cases and other rare events are most relevant and worth learning. In such occasions, data augmentation techniques offer enormous benefits to alleviate the issue of limited training data. However, unlike in image data, this situation becomes challenging when dealing with sequential one-dimensional data, where straightforward methods for data augmentation fail to fully capture the properties of underlying processes. Our proposed research will investigate the most recent advances in artificial intelligence with the objective of building a multi-purpose time-series data augmentation engine. This project is expected to develop new or improved augmentation techniques that will ease the data requirements of existing methods and expand their applicability to a wide range of applications.

Kyle Dunphy
Faculty Supervisor: 
Ayan Sadhu
Partner University: