Primitives for Online Time Series Analysis

New York University, New York NY

Investigators

Abstract

This project aims to build a foundational library of primitives for calculating statistics on time series at online or near-online speeds. A time series is simply a collection of data arriving in time order. They arise in fields ranging from physics to finance or medicine or music. Often the data comes from sensors, that are becoming deployed in rapidly increasing number of applications and whose data rates are continuing to increase sharply. Fast (online) response is desirable in many applications (e.g., remote aiming of a telescope to a fast-moving object of interest). The faster the volumes of the time series data grow, the more important it is to derive summary statistics about the data. This project aims to provide tools for the rapid incremental computation of basic statistics (such as correlation and burst detection) on collections of thousands or more time series. This work has two thrusts: the development of algorithmically efficient methods for calculating such statistics and the engineering of these methods and already known approaches to yield practical implementations. To ensure concreteness, the research and implementation activities are driven in part by collaborations with physicists and other scientists. The statistical primitives will enhance research and the national cyberinfrastructure in any application requiring the fusion of time series information. Note that this work encompasses all kinds of time series, including time series for music and other media. This may attract K-12 students, drawing them to science through applications such as query by humming. The project's Web site (http://cs.nyu.edu/cs/faculty/shasha/papers/statstream.html) is used to broaden dissemination of results.