Large-Margin Classifier Selection via Decision Boundary Stability

• Activity Number: 413
• Type: Topic Contributed
• Date/Time: Tuesday, August 6, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Wei Sun*+ and Guang Cheng and Yufeng Liu
• Companies: Purdue University and Purdue University and The University of North Carolina
• Keywords: Asymptotic normality ; large-margin classifier ; model selection ; selection consistency ; stability

Abstract

Large-margin methods have been widely used in classification. However, which classifier should be chosen in practice remains to be an open question. The existing criterion for classifier comparison is on generalization error, while it does not consider the variation in prediction. In this article we first show that the difference of generalization errors induced from two classifiers may not be significant. Considering the variation in prediction, we introduce a new concept of decision boundary stability to measure the variability of classifiers. To incorporate both accuracy and stability, we propose a two-stage classifier selection algorithm. First, the potentially good classifiers are initially selected based on the statistical testing of generalization errors; Second, the optimal classifier is then chosen as the one with the largest decision boundary stability. The proposed two-stage selection algorithm is shown to be selection consistent in that the selected classifier asymptotically obtains both minimal generalization error and maximal decision boundary stability. Its effectiveness is illustrated in various simulated examples and real data sets.