Evaluation of machine learning approaches for change-proneness prediction using code smells

Abstract

In the field of technology, software is an essential driver of business and industry. Software undergoes changes due to maintenance activities initiated by bug fixing, improved documentation, and new requirements of users. In software, code smells are indicators of a system which may give maintenance problem in future. This paper evaluates six types of machine learning algorithms to predict change-proneness using code smells as predictors for various versions of four Java-coded applications. Two approaches are used: method 1-random undersampling is done before Feature selection; method 2-feature selection is done prior to random undersampling. This paper concludes that gene expression programming (GEP) gives maximum AUC value, whereas cascade correlation network (CCR), treeboost, and PNN\GRNN algorithms are among top algorithms to predict F-measure, precision, recall, and accuracy. Also, GOD and L_M code smells are good predictors of software change-proneness. Results show that method 1 outperforms method 2.