Data pre-processing by genetic algorithms for bankruptcy prediction

摘要

Bankruptcy prediction has been approached by data mining techniques. However, since data pre-processing including feature selection or dimensionality reduction and data reduction is a very important stage for successful data mining, very few consider performing both tasks to examine the impact of data pre-processing on prediction performance. This paper applies genetic algorithms, which have been widely used for the data pre-processing tasks, for feature selection and data reduction over a public bankruptcy prediction dataset. In particular, the experiments based on different priorities of performing feature selection and data reduction are conducted. The results show that performing data reduction only can allow the support vector machine (SVM) classifier to provide the highest rate of prediction accuracy. However, executing both feature selection and data reduction with different priorities performs the same. They not only largely reduce the dataset size, but also keep the similar performance as SVM without data pre-processing.