Training a Classifier for Optimal Classification Error

• Activity Number: 251
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Frans H.J. Kanfer*+ and Ryno Potgieter and Sollie Millard
• Companies: University of Pretoria and University of Pretoria and University of Pretoria
• Keywords: Classification ; prescribed misclassification rate ; Sequential training ; limited training cases

Abstract

Training a classifier to a pre-determined level can be achieved by prescribing the misclassification rate and the certainty of obtaining such a rate. A sequential training procedure has been introduced in literature training a selected classifier to such requirements, with an additional advantage of limiting required training cases. A shortfall of the procedure is that it does not accounting for unfeasible specifications. This paper presents a sequential procedure which follows an approach of estimating the best feasible misclassification rate at a prescribed level of accuracy. It can also be applied to any classification algorithm and limit the training cases as required by the prescribed specification. Simulation results are presented for LDA and KNN classification. A micro array data application is also discussed.