Group and Within-Group Variable Selection via Convex Penalties

• Activity Number: 532
• Type: Contributed
• Date/Time: Wednesday, August 4, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Yun Li*+
• Companies: University of Michigan
• Keywords: variable selection ; convex penalties ; within group

Abstract

In many scientific applications, there is a natural grouping of prediction variables. Such as, genes can be grouped by biological pathways. Traditional variable selection methods make selection based only on the strength of individual variables rather than on the strengths of groups. Existing group variable selection methods have an ``all-in-all-out'' limitation. In many important real problems, however, we may want to keep the flexibility of selecting variables within a group. We may want to not only remove unimportant groups effectively, but also identify important variables within important groups. We propose a new group variable selection method that not only removes unimportant groups effectively, but also keeps the flexibility of selecting variables within a group. Unlike earlier methods addressing the same problem, the new method utilizes a convex, rather than concave, criterion.