UC Davis

Given a set of time series data, the goal for change point detection is to locate, if any, those time points at which some characteristic of the data changes. The need for change point detection arises in many contexts, including stock market prediction, weather forecast, and air pollution monitoring. This thesis studies change point detection for three types of complex data: time series of astronomical images, sequences of undirected graphs, and time-evolving dynamic networks. From each of these three problems, the minimum description length principle is invoked to derive a model selection criterion, which is shown to yield statistically consistent estimates for the change points and other model parameters. Practical, tailored algorithms are also developed to compute these estimates.