سمينار علمي

Bayesian Biclustering with Variable and Subject Selection

دكتر وحيد پرتوي نيا

دانشگاه پلي‌تكنيك مونترال كانادا

مكان: سالن خوارزمی دانشکده علوم رياضي

زمان:يكشنبه 19/5/1393 ساعت 10 صبح

This talk briefly introduces the Bayesian view to the clustering problem.Variable selection also is an issue in clustering high-dimensional data, but hard to deal with, especially in clustering since there is no clear response variable to guide the search. I propose a parametric hierarchical Bayesian model and show how variable importances can be quantified for classification and clustering using the introduced model in a coherent way.  Another way to handle high-dimensionality in clustering is to group similar variables simultaneously, i.e. Biclustering. I suggest a fully automatic procedure under which hyper-parameters of the model can be estimated by empirical Bayes. The statistical package related to clustering with variable selection, "bclust", is published on R-CRAN, and the package related to biclustering method, "baybi", is under development on R-Forge.