The effects of feature selection and model selection on the correctness of classification

In this research we took an experiment of two feature selection methods - eta square and stepwise methods on two classification models - back propagation neural network (BPNN) and general regression neural network (GRNN) to study the effects on the correctness of firm bankruptcy classification. The correctness includes the average classification correctness and the power of bankruptcy classification which is the probability we conclude failure if firms are in crisis. The data were sampled from the listed electronic companies in Taiwan´s market from 1999 to 2006. The experimental reports showed that feature selection has more influences on average correctness and power than model selection. Overall, the stepwise method has the highest correctness among these four combinations which are the two feature selections and two model selections but the two models BRNN and GRNN has not much difference in our experiment.