Boosting neural networks in real world applications: An empirical study

Abstract

Boosting techniques allow the combination of a collection of sequentially trained neural networks into an ensemble whose classification performance is superior to any of the individual neural networks. Empirical studies on the performance of boosting neural networks in optical character recognition have demonstrated significant improvements in classification. In this paper we report on an empirical study of boosting neural networks for classifying business data from real world databases. These data often contain noise and subjective or even contradictory classifications. Therefore, classification of such business data is a hard problem in practical applications. Two boosting algorithms were tested in this empirical study. The experimental results have shown that boosting neural networks indeed improved the classification performance. With one data set, we have achieved to date the best classification result, which had never been achieved using single and committee neural networks.