ATD: Inductive Spatiotemporal Graph Encoding for Interpretable and Transferable Deep Learning with Application in Human Dynamics

This project aims to develop novel statistical and computational methods to address the emerging issues in human dynamics, including detecting anomalies in human mobility, predicting potential disastrous damage, and monitoring disease outbreak in our society. At present, the proliferation of digital mobility data, such as phone records, GPS traces, and social media posts, combined with the outstanding predictive power of statistical learning, triggered the applications of deep learning by using the mobility of individuals as a proxy to study human dynamics. This project will provide interpretable and transferable methods for discovering unusual events in any super-large spatiotemporal data set, inspire a new line of research in big data analytics, and offer a unique opportunity for students to participate in cutting-edge and interdisciplinary big data research. This project will support one graduate student per year at each university for each of the three years of the project.

This project uses a proxy such as mobile phone dynamic graphs to develop human dynamics models that can be used for threat detection and infectious disease prediction. Mobility data is naturally represented as a dynamic graph, where any individual node represents a location or a group of people, and its connections correspond to measures of mobility between the nodes. The anomaly node or connection can be used for detecting threats or disasters. One concrete application is predicting the super-spreaders and the number of infections of COVID-19 using the SafeGraph data that consists of the trajectory of millions of mobile phone users, which is clinically essential to harness the virus spreading. Questions to be explored in this research project include: (1) How to encode the dynamic graphs evolving spatiotemporal information at node (or subgraph) level into low-dimensional embedding vectors that can be used as feature inputs for further downstream prediction and inference (2) How to quantify the importance of nodes and connections in the dynamic graph for virus transmission (3) How to transfer knowledge from mobility data and multi-source data of the source locations to predict the epidemic trends for new locations with fewer observations. This project will address these questions and develop a general framework for inductive spatial encoding in large dynamic graphs, which enables interpretable, and transferable learning for different locations or tasks. The fast, transferable computing principles developed in dynamic graph modeling are fundamental and indispensable tools for "big data" computation and autonomous systems. The principles will be widely applicable to diverse fields of sciences, engineering, and humanities.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.