Deep neural network based hybrid approach for software defect prediction using software metrics

Abstract

In the field of early prediction of software defects, various techniques have been developed such as data mining techniques, machine learning techniques. Still early prediction of defects is a challenging task which needs to be addressed and can be improved by getting higher classification rate of defect prediction. With the aim of addressing this issue, we introduce a hybrid approach by combining genetic algorithm (GA) for feature optimization with deep neural network (DNN) for classification. An improved version of GA is incorporated which includes a new technique for chromosome designing and fitness function computation. DNN technique is also improvised using adaptive auto-encoder which provides better representation of selected software features. The improved efficiency of the proposed hybrid approach due to deployment of optimization technique is demonstrated through case studies. An experimental study is carried out for software defect prediction by considering PROMISE dataset using MATLAB tool. In this study, we have used the proposed novel method for classification and defect prediction. Comparative study shows that the proposed approach of prediction of software defects performs better when compared with other techniques where 97.82% accuracy is obtained for KC1 dataset, 97.59% accuracy is obtained for CM1 dataset, 97.96% accuracy is obtained for PC3 dataset and 98.00% accuracy is obtained for PC4 dataset.