Title: Dynamic Clustering Methods for Analyzing Complex Systems
Advisor: Professor George Michailidis
Committee Members: Associate Professor Kerby Shedden, Associate Professor Lada, Adamic (School of Information)
Abstract: Driven by advances in computer technology, businesses and organizations are increasingly able to collect streams of data that capture the behavior of components in complex systems. Suppose we observe noisy snapshots, at the component level, of a complex system over time. In particular, we measure the same covariates (behavioral characteristics, features) for each component at each point in time. A challenging issue is to identify and interpret significant component behaviors and interactions. Further, in many cases, it is impractical or even impossible, to store and organize the full data for efficient retrieval and analysis. Thus, algorithms that do not require the full data and operate given the input of information up to a point in time are essential to the analysis of large systems. Systems with such characteristics are encountered in biology, business, economics, transportation, among others. To gain insight into these interactions we develop a dynamic online biclustering method based on the plaid model. We introduce a regularization framework that smooths the model parameters over time, thus allowing us to identify persistent groups (components) and the critical features that separate them consistently over time. An application to data obtained from an electronic trading market is used to illustrate this methodology. Extensions that incorporate knowledge of component interaction and system geometry through graph regularization are discussed.