University of Montreal
A Bayesian lasso functional model-based clustering model
We develop a flexible model for the analysis and clustering of complete or sparse time-course data.
The model combines functional analysis and model-based clustering. The functional framework is used to model time-course data. Principal functional components are described by score coefficients which embed the curves in a much lower-dimensional space. Model based clustering is performed on the score space, thus avoiding the curse of dimensionality in the curves' space. The model is embedded into a Bayesian framework. We first develop an approximation of the marginal log-likelihood MLL that allows us to perform a MLL based model selection. We then developed a Bayesian version of the lasso penalty in order to render the model selection step more efficient. The number of clusters as well as the dimension of the score space are determined via this Bayesian lasso penalty model. Monte Carlo techniques are used in order to estimate the normalizing constants for different values of the penalty parameters. One of the advantages of the Bayesian lasso model is that it avoids the need to perform a costly cross-validation to select the penalty parameters. We show some applications to the analysis of gene-expression data.
This is joint work with Folly Adjogou, Université de Montréal, Wolfgang Raffelsberger, Université de Strasbourg, and Karin Dorman, Iowa State University.
Refreshments at 3:45pm in Snedecor 2101.