CRII: III: Multiresolution Tensor Learning for Scalable and Interpretable Spatiotemporal Analysis

点击翻译按钮获取中文摘要

A Framework for Scalable and Interpretable Spatiotemporal Data AnalysisThe past few decades have witnessed an explosion of large-scale spatiotemporal data. At the same time, domain experts and policy makers need data analysis tools that are explainable in order to trust and deploy them. However, analyzing spatiotemporal data is highly challenging, as such data often demonstrates complex correlations, high dimensionality, and contains multiple spatiotemporal resolutions. Meeting this challenge will require new methods that can handle highly correlated data, generalize to higher dimensions, and reason at different granularities. To address these challenges, this project will develop a spatiotemporal data analysis framework that is both scalable and interpretable. The project will apply this framework to tackle challenging problems in climate informatics, sports analytics, and intelligent transportation. This project will establish a large-scale spatiotemporal data repository and distribute open source software to benchmarkresearch progress. The educational component will develop new courses geared towards the intersection of machine learning and spatiotemporal analysis. Additionally, the principal investigator plans to continue outreach activities that involve giving tutorials, organizing workshops at relevant conferences.The long term goal of this project is to improve the scalability and interpretability of spatiotemporal analysis tools. The principal investigator has initiated a framework of tensor learning that can capture higher-order correlations and address high-dimensional issues. This project will significantly expand the framework to exploit the multiresolution nature of spatiotemporal data. In particular, the principal investigator will develop a multiresolution tensor learning framework and integrate this framework into existing models, including latent factor models, Gaussian processes and graph neural networks. This approach leverages fast tensor optimization algorithms to recognize spatiotemporal patterns at multiple granularities. This project will lead to novel techniques to automatically discover latent semantics, quantify uncertainty, and learn feature representations from spatiotemporal data in a scalable and interpretable fashion. It will further contribute to our burgeoning understanding of advanced optimization and statistical tools such as multigrid optimization, tree-based Gaussian processes and geometric deep learning.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.