Classification of Sparse and Irregularly-Sampled Time Series Data

22 Jun
Thursday, 06/22/2017 9:30am to 11:30am
Computer Science Building, Room 151
Ph.D. Dissertation Proposal Defense
Speaker: Steve Li

"Classification of Sparse and Irregularly-Sampled Time Series Data"

Sparse and irregularly-sampled time series data occur in various application areas including climate science, ecology, biology, and medicine. In this work, we study models and methods of classifying data instances that are sparse and irregularly-sampled time series. The properties of such data can result in substantial uncertainty about the values of the underlying temporal processes, while making the data difficult to deal with using standard classification methods that assume fixed-dimensional feature spaces. To address these challenges, we introduce a framework based on Gaussian process (GP) regression that represents each time series through an induced posterior GP. We then propose an uncertainty-aware kernel over the Gaussian process representation that can be used for time series classification. We show how to accelerate the proposed kernel computations using specialized random Fourier features and Fastfood approximation. Next, we present an algorithm to scale up the computation of GP posteriors based on structured kernel interpolation and the Lanczos approximation method. We show how to combine this GP module with arbitrary black-box classifiers and efficiently train the entire model discriminatively end-to-end. Finally, we point out that basic GP-based time series models ignore the complex global structure shared across the whole data set. We propose new methods to learn from incomplete data based on recent advances in deep generative models, and apply these ideas to learn from sequential data including irregularly-sampled time series.

Advisor: Benjamin Marlin