Model selection in change-point problems

Dr Haeran Cho, University of Bristol (part of the Statistics Seminar Series)

Lately, there has been a surge of interest in research for computationally fast and statistically efficient methods for change-point detection, as nonstationarities frequently observed in real-life datasets are often attributed to structural breaks in the underlying stochastic properties. In multiple change-point detection, model selection via estimating the total number of change-points poses as a challenge, particularly when the dimensionality of the data is large. In this talk, I will address the model selection in change-point problems in two different settings: when $p = 1$ where one can benefit from localised application of an information criterion, and when $p$ is large where the change-point detection problem can be translated to that of detecting pervasive and latent 'factors'.