|
Abstract:
|
We present a novel model selection methodology, based on the well-known principle of bit-allocation in rate-distortion theory. The method turns the combinatorial subset selection problem, across clusters and variables, into a simple line search. Moreover, in the special case where data objects form their own clusters, the methodology generalizes to subset selection in a multiple testing framework. The selection method is applied to a time course gene expression dataset of two proliferating stem cell lines. Our analysis provides sparse and easy-to-interpret gene cluster models. In addition, by incorporating subset selection into testing, the power of detection is significantly increased. The method provides sparse model representations for each gene. Thus, we find that the significant list is dominated by main cell line and time effects, and a few informative interaction models.
|
