Stability Selection in Dimension Reduction

• Activity Number: 356
• Type: Contributed
• Date/Time: Tuesday, July 31, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Wenbo Wu*+
• Keywords: stability selection ; re-sampling ; shrinkage estimation ; sufficient dimension reduction

Abstract

Many existing methods, such as sparse eigen-decomposition estimation (SED) and sparse sufficient dimension reduction (SSDR) transformed a dimension reduction problem into a regression-type formulation and adopted shrinkage estimation procedures to produce sparse and accurate solutions. But such shrinkage estimations can sometimes be inconsistent without choosing the effective tuning parameters. We propose a stability selection method which is based on sub-sampling in combining with random weights selection to widen the effective tuning parameter range which will hence lead to a more consistent estimation procedure. Stability selection can be directly incorporated with SED and SSDR procedure to provide more accurate and more consistent estimation of the dimension and directions in central subspace. Simulation results demonstrate the effectiveness of stability selection that used in dimension reduction.