A Comparison of Classification Methods for Robustness

• Activity Number: 497
• Type: Topic Contributed
• Date/Time: Wednesday, August 1, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Yoonkyung Lee*+ and Rui Wang
• Companies: The Ohio State University and The Ohio State University
• Keywords: classification ; robustness ; excess error ; mislabeling

Abstract

Classification aims at accurate prediction of categorical outcomes, and it may not require a formal model for the underlying data generation mechanism. However, as with procedures for statistical modeling, robustness is desirable for classification. Two pertinent situations are when there is discrepancy between the optimal classification boundary and its best approximation by a classification method given a class of discriminant functions, and when there are errors in the class labels. We compare robustness of various classification methods including both model-based and algorithmic methods in the two settings. Comparisons are made on the basis of excess error rates as the approximation error varies in the first setting and the proportion of mislabeled data varies in the second setting.