Georgia Institute of Technology

Abstract: Change-point detection is a classic statistical framework for detecting a change in the distribution of a sequence of data. In this talk, I will focus on its connection with machine learning and anomaly detection, and illustrate by our two recent work along this direction. While classic change-point detection usually assumes i.i.d. data and parametric forms of the data distributions, when dealing with machine learning problems we may need to go beyond these settings. The first work considers detecting a change in a network where one observes a sequence of correlated discrete events on the nodes. The second work presents a distribution-free kernel based method leveraging minimum mean discrepancy (MMD) statistic. The common themes are to construct detection statistics that are suitable for machine learning tasks and to control the false alarm rate via a powerful change-of-measure technique. This is a joint work with Shuang Li, Le Song, Mehrdad Farajtba and Apart Verma.

Bio: Yao Xie is an Assistant Professor in the H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology. She received her Ph.D. in Electrical Engineering (minor in Mathematics) from Stanford University in 2011. Prior joining Georgia Tech, she worked as a Research Scientist at Duke University. Her research areas include computational statistics, signal processing, and machine learning, in providing theoretical insights, developing computationally efficient and statistically powerful algorithms for various application, including sensor networks, social networks, imaging, material science, geophysics, communications. She received a Best Student Paper Award at Annual Asilomar Conference on Signals, Systems and Computers in 2005, Finalist of Best Student Paper Award in ICASSP Conference in 2007, and the National Science Foundation (NSF) CAREER Award in 2017.