Performance Prediction for Classification Systems

Abstract

Performance prediction for classification systems is important. We present new techniques for such predictions in settings where data items are to be classified into two categories. Our results can be integrated into existing classification systems and provide an accurate and predictable tool for data mining. In any given classification case, our approach uses all available training data for building the classification scheme and guarantees zero classification errors on the training data. We re-use the same training data to predict the performance of that scheme. The method proposed here enables control of errors over two types of error for the classification task.

This work was done when the author studied in the Computer Science Department of the University of Texas at Dallas. The author is currently working for Alcatel Network Systems.