[Mobilizeplans-starstudents] TODAY - Mobilize Center Special Seminar: Exploiting Patterns for Efficient and Accurate Time-Series Analysis

Joy P. Ku joyku at stanford.edu
Thu May 18 07:06:40 PDT 2017

We invite you to join us TODAY, Thursday, May 18th, for a special Mobilize Center Seminar featuring postdoctoral candidate, John Paparrizos. Note the alternate time and room.

TITLE: Exploiting Patterns for Efficient and Accurate Time-Series Analysis

SPEAKER: John Paparrizos, PhD Candidate in Computer Science, Columbia University

WHEN: TODAY - Thursday, May 18th, at 10:00am

WHERE: Gates 415

Large volumes of time series appear in almost every discipline, including astronomy, biology, meteorology, medicine, finance, robotics, engineering, and others. This proliferation of continuously evolving data has generated a substantial interest in the analysis and mining of time series. In particular, clustering and classification methods for time series have received significant attention, not only as powerful stand-alone methods, but also as important subroutines and building blocks in other tasks or systems. The detection and extraction of patterns, despite distortions that are characteristic of the time series, are crucial steps to devise effective methods for real-world problems. In this talk, I present novel algorithms for fast and accurate clustering and classification of time series and I analyze the performance of pattern extraction in time series over two important real-world applications. Specifically, by exploiting the correlation of the shapes of the time series, I develop a scalable and accurate centroid-based clustering algorithm. Having the ability to select (through clustering) a small number of prototypes that effectively summarize the underlying time series, I present a prototype-based sparse coding mechanism. This mechanism relies on principles that permit designing a computationally tractable adjacency matrix along with matrix's eigen-decomposition. Additionally, as outliers in the datasets and noise in time series can significantly affect the accuracy of clustering methods, I leverage patterns from subsequences of the time series and develop an outlier-aware clustering algorithm. I further explore how these clustering methods can lead to efficient and accurate techniques for nearest-neighbor classification algorithms. Finally, I show the impact of pattern extraction from time-varying measurements in two real-world applications: (i) prediction of the impact of scientific concepts and articles through temporal analysis of characteristics extracted from metadata and full text of scientific articles; and (iii) detection of devastating diseases in web search engine logs to alert users of potential health risks.

Joy P. Ku, PhD
Project Manager, SimTK<http://simtk.org/>
Director of Communications & Training, NCSRR<http://opensim.stanford.edu/>
Director of Communications & Engagement, Mobilize Center<http://mobilize.stanford.edu/>
Stanford University
(w)  650.736.8434, (f) 650.723.7461
Email:  joyku at stanford.edu

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://simtk.org/pipermail/mobilizeplans-starstudents/attachments/20170518/add066ba/attachment.html>

More information about the Mobilizeplans-starstudents mailing list