Analysis of evolutionary algorithms to improve software defect prediction

Defect prediction of software is necessary to determine defective parts of software. Defect prediction models are elaborated with the help of software metrics when combined with defective data to predict the classes that are defective. In this paper we have used datasets that statistically resolve the relationship among software metrics and defect vulnerability. The main intent of this paper are 1) Feature selection for defect prediction using proposed evolutionary algorithm 2) Comparing machine learning techniques 3) Use of precision and recall as performance measure for defect prediction 4) 10- fold validation is performed on every model. In this discourse, we predict defective class using 5 machine learning techniques and 2 evolutionary techniques for feature selection. In this work, we have applied evolutionary algorithms for feature selection suitable for each of the classification techniques applied on five open source android packages. Finally, for validation of calculated results, 10-fold validation is used. The results show that using evolutionary algorithms for feature selection can improve precision and recall for RF, DT and SVM. Precision and recall have best rise using SVM model. The use of evolutionary algorithms don't effect precision and recall for statistical classifier. The results obtained from evaluation thus confirm about the prediction of default classes using evolutionary algorithms is better than using only machine learning techniques. The analyzed and calculated results gave us the view about the usage of evolutionary algorithm with statistical classifier is of no use.