A Product Partition Model for Detecting Change Points on General Graphs

• Activity Number: 134
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Computing
• Author(s): Xiaofei Wang*+ and John W. Emerson
• Companies: Yale and Yale University
• Keywords: Bayesian ; change point ; product partition model ; image segmentation ; graph

Abstract

The product partition model (PPM) has been proposed as a Bayesian solution to the traditional change point problem of estimating the location of mean shifts in time series data. Subsequent extensions of the model have then expanded the PPM to address two-dimensional change points on a rectangular lattice. In this presentation, we present a new PPM to handle change points on a general graph. We show that our PPM is a generalization of the two-dimensional lattice PPM and very closely related to the one-dimensional time series PPM. As with traditional PPMs, our algorithm employs MCMC sampling to produce a posterior distribution of partitions. For the two-dimensional rectangular lattice case, we show via simulation that our method performs favorably (generally faster runtime and higher robustness to underlying tuning parameters) compared to the existing PPM algorithm devised for this special case. Finally, we will show an application of our method to image segmentation.