A0715
Title: A random projection based technique for change point detection in high-dimension
Authors: Nilabja Guha - University of Manchester (United Kingdom) [presenting]
Jyotishka Datta - Virginia Polytechnic Institute and State University (United States)
Abstract: In many applications, such as economics, social science, and finance, changes in data-generating distributions are observed with time. The observed variable may depend on covariates through a mean structure, where the mean structure may change with time. There can also be changes in the underlying covariance structure. A Bayesian framework of change point estimation is presented for high-dimensional observations. Such high-dimensional observations may appear in many practical applications where the high-dimensional mean parameter or the covariance structure changes with time, such as high-frequency financial data. A lower-dimensional embedding is presented based on random projection. Change point estimation consistency and convergence rate are established even when the dimension of the observations can be much larger than the number of observations.
