Simple Tiered Classifiers

• Activity Number: 325
• Type: Invited
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: IMS
• Author(s): Peter Gavin Hall*+ and Jinghao Xue and Yingcun Xia
• Companies: University of Melbourne and University College London and National University of Singapore
• Keywords: Classification

Abstract

In some classification problems, notably those that involve class-unbalanced data, at least one of the populations, often the less common one, seems to defy correct classification using conventional methods that are effective for other data. In this talk we take up the issue of difficult-to-classify populations or subpopulations, suggesting methodology that implements classification in at least two tiers, thereby constructing a more sophisticated classifier to deal with relatively complex data structures. Empirical studies on real and simulated data show that three two-tier classifiers, which are respectively extensions of linear discriminant analysis, linear logistic regression and support vector machines, can reduce noticeably the relatively high misclassification error of their original single-tier counterparts, without increasing much the computational complexity of the latter.