Model-Free Sure Screening via Maximum Correlation

• Activity Number: 80
• Type: Contributed
• Date/Time: Sunday, August 3, 2014 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Qiming Huang*+ and Yu Zhu
• Companies: and Purdue University
• Keywords: B-spline ; Maximum correlation ; Optimal transformation ; Sure screening property.

Abstract

We consider the problem of screening features in an ultrahigh-dimensional setting. Based on maximum correlation, we develop a new procedure called MC-SIS for sure independence feature screening. Compared with other sure independence screening methods, MC-SIS is free of model and distributional assumptions. We further establish the sure screening property of MC-SIS. Simulation examples and a real data application are used to demonstrate the performance of MC-SIS as well as compare MC-SIS with other existing sure screening methods. The comparison results show that MC-SIS can significantly outperform other methods when the model assumptions of these methods are violated, and it remains competitive when the model assumptions hold.