III: Small: Robust Algorithms for Multi-Task Learning of Spatio-Temporal Data

Michigan State University, East Lansing MI

Investigators

Abstract

Recent years have witnessed an explosive growth of data collected from different geographical locations at different time points, known as spatio-temporal data. Spatio-temporal data are generated from a wide array of sensing technology and through scientific simulations. Analysis of the big spatio-temporal data is crucial as it supports key applications of national importance such as environmental sustainability, healthcare, and national security. However, in order to realize its full benefits, there are major analytical and computational challenges that must be overcome. This interdisciplinary research seeks to overcome some of the challenges by developing innovative prediction algorithms. As proof of concept, the algorithms will be incorporated into an ensemble prediction system developed by one of the principal investigators to enhance the nation's seasonal drought forecast capability. Benchmark data sets will be created and made available to other researchers for evaluating their algorithms. The project web page (http://www.cse.msu.edu/~ptan/project/mtl) will be used to disseminate experimental results, published papers, and software developed in this project. The graduate students who participate in the project will be trained to conduct cutting edge multidisciplinary research as next-generation data scientists. There are many spatio-temporal prediction tasks that involve solving multiple related sub-problems. Rather than designing the prediction model for each sub-problem independently, it would be desirable to solve the prediction tasks jointly by exploiting their spatio-temporal autocorrelations, thus making it a natural fit for applying a multi-task learning paradigm. This proposal centers around the following four key contributions to address the challenges of applying multi-task learning to large-scale spatio-temporal data. First, a space-efficient online multi-task learning algorithm, with theoretically proven convergence rates, will be developed to handle the massive volume of data. Second, a deep learning framework will be developed to extract salient spatio-temporal features for multi-task predictions. Third, novel multi-task learning algorithms will be developed to deal with multi-scale variables, which are becoming increasingly prevalent with the proliferation of high-resolution sensors and simulation data generated at increasingly finer resolutions. Finally, the proposed algorithms will be applied to various spatio-temporal domains, including climate and environmental sciences. Overall the proposed research will advance current state-of-the-art by developing innovative prediction algorithms that consider the inherent spatial and temporal variability of the data and provide solutions with high predictive accuracy and learning efficiency.