Spatiotemporal forecasting is critical for various domains such as highway traffic and weather forecasting. Capturing both spatial and temporal dependencies are the key factors for accurate forecasting. I will present my research on graph neural networks for large scale spatiotemporal forecasting. Training graph neural network models for large scale highway is challenging due to the GPU memory capacity and computational bottlenecks.  I will present a new graph partitioning based diffusion convolution recurrent neural network (DCRNN) in which a large graph is partitioned into several subgraphs, each of which is trained simultaneously on multiple GPUs [1]. Next, I will present a transfer learning approach for DCRNN, wherein a model trained on one region of the highway network can be used to forecast traffic on unseen region of the highway network [2]. Moreover, I will discuss how the DCRNN can be used to forecast data traffic in Esnet, the DOE’s dedicated science network, where the traffic patterns are dynamic and do not exhibit long-term spatial and temporal regularity [3]. Finally, I will give a high-level overview on a domain-informed natural language processing (NLP) pipeline developed exclusively for analyzing the impact of COVID-19 on the critical infrastructure such as government and supply chain sectors. The pipeline includes active learning approach to label most informative documents, unsupervised topic modeling, and semantic search using pretrained language models.

[1] Mallick T, Balaprakash P, Rask E, Macfarlane J. Graph-partitioning-based diffusion convolutional recurrent neural network for large-scale traffic forecasting. Transportation Research Record. 2020; 2674(9), pp.473-488.[2] Mallick T, Balaprakash P, Rask E, Macfarlane J. Transfer learning with graph neural networks for short-term highway traffic forecasting. In 2020 25th International Conference on Pattern Recognition (ICPR). 2021; pp.10367-10374.[3] Mallick T, Kiran M, Mohammed B and Balaprakash P. Dynamic Graph Neural Network for Traffic Forecasting in Wide Area Networks. In 2020 IEEE International Conference on Big Data (Big Data). 2020; pp. 1-10.

Zoom Link: https://argonne.zoomgov.com/j/1619592864