Transferability of Graph Neural Networks for Time Series Applications

Abstract

Transfer learning enabled machine learning tasks with scarce data to achieve superhuman performance in multiple domains like computer vision and natural language processing. However, knowledge transfer's success was mostly on grid structured data and using convolutional neural networks that assume local, hierarchical, and stationary data. Time series data in several applications, specifically doesn't meet these assumptions. This renders traditional transfer learning irrelevant with the potential leading to negative transfer. After achieving superior performance on high-dimensional data like social networks and recommender systems, graph neural networks are currently applied to time series data. In this thesis, we investigate the transferability of graph neural networks on time series data compared to traditional time series algorithms. We also explore a new graph similarity approach and compare its effect on time series algorithms pretraining and negative transfer for pandemic time series forecasting.