Time:16:00 pm, Dec. 17th
Lecturer: Dr. ZHOU Wang, professor of National University of Singapore
Content: Information criterion is very important in model selection and variable selection, more so in high dimensional settings. There are several existing ones designed in high dimension settings like high dimensional Bayesian Information Criteria(HBIC) and extended Bayesian Information Criterion(EBIC), which are proved to be useful in both theory and application. However, the subtle balance between unknown parameters and the complexity of the model is worthy to be further studied. In the paper, we propose a new Bayesian information criterion, which allows the dimensionality of covariates to grow exponentially fast with the sample size. Model selection consistency for both unpenalized and penalized estimators are established. Extensive simulation studies in commonly used models show that our information criterion has substantial improvement against other major competitors. Thus we name our method as improved Bayesian Information Criterion(IBIC). Moreover, we extend IBIC to select thresholding paremeter for sparse covariance matrix estimation and the results are promising.
Interested teachers and students are all welcomed!