Evolutionary feature selection in boosting

Abstract

The purpose of this study is to clarify the effectiveness of a new type of weak learner in boosting for pattern classification. Our weak learner is called EFS (evolutionary feature selection). The EFS has two aspects: the first is a feature-subset selector for pattern classification. The EFS selects effective combinations of features using an evolutionary technique. An entropy-based criterion called VQCCE (vector-quantized conditional class entropy) is used for the evaluation of feature-combinations. The second is a weak learner in boosting. We utilize the vector-quantization in the EFS as the weak learner. In this paper, we apply our method to some benchmark problems and discuss the effectiveness of our method, in comparing with a conventional boosting with C4.5 decision trees