29th November 2005 — James Griffin: Dependent Dirichlet Processes
James Griffin is giving a guest talk:
Bayesian Nonparametric Inference in Time Series
The Dirichlet process is a popular tool for Bayesian nonparametric analysis. It has recently been used to build clustering algorithms with an unknown number of components (Escobar and West). The data are assume to be identically distributed and drawn independently from a mixture distribution. Each component of the mixture distribution defines a cluster. If the data are observed over time, the data cannot be assumed identically distributed. This talk will discuss nonparametric priors. This approach leads to algorithm that allow the allocation of observations to cluster to depend on time. For example, Zhu et al (2005) use a similar approach to cluster emails allowing for their time of arrival. The talk will build on the work of Griffin and Steel (2006).
- Bayesian density estimation and inference using mixtures, Michael D. Escobar and Mike West, Journal of the American Statistical Association, 90(430):577–588, 1995.
- Order-based dependent Dirichlet processes, Griffin, J. E. and M. F. J. Steel, Journal of the American Statistical Association (forthcoming), 2006.
- Time-sensitive Dirichlet process mixture models, Zhu, X., Z. Ghahramani and J. Lafferty, Technical report: CMU-CALD-05-104, School of Computer Science, Carnegie Mellon University, 2005.