Fast and Interpretable Tensor Methods for Spatiotemporal Analysis

Multivariate spatiotemporal data is ubiquitous in science and engineering, from sports analytics to neuroscience. Such data can be naturally represented as a multiway tensor. Tensor latent factor models provide a powerful tool for reducing the dimensionality and discovering the higher-order latent structures from data. However, existing tensor models are often slow or fail to yield latent factors that are easy to interpret by domain experts. In this talk, I will demonstrate advances in tensor methods to generate interpretable latent factors for high-dimensional spatiotemporal data. In particular, I will discuss (1) a multiresolution tensor learning algorithm, that can leverage the multicale property of high resolution spatial data, to speed up training and learn interpretable patterns. (2) a tensor latent feature learning algorithm that can learn binary representations of data that are both memory efficient and easy to interpret. We provide theoretical guarantees for our optimization algorithms and demonstrate their applications to real-world data from basketball plays and neuroscience.