Speaker: Xin Gao (York University)

Title: Integrative feature selection on correlated platforms with ultra high dimensionality

Date: November 20, 2014

Time: 2:30 pm

Location: Ottawa University, KED 585 – Room 2015

ABSTRACT: In this article, we consider the problem of perform feature estimation and selection when the features manifest in different parameters in different experimental platforms. To integrate the data across different experiment platforms, composite likelihood offers a flexible tool to compound the marginal densities over correlated data sets. We propose a unified framework for feature selection and model selection when the data are integrated from different sources. The framework provides two strategies: feature selection through grouped penalty function on composite likelihood so that same feature can be selected based on information across different experimental sources; model selection based on composite likelihood information criterion so that information can be integrated over different platforms. We establish the asymptotic ORACLE property for the composite likelihood estimates with group penalty in ultra high dimensional setting. Under certain exponential moment conditions, the proposed composite likelihood is selection consistent even when the true model size is divergent.