Analyzing and improving information gain of metrics used in software defect prediction in decision trees

Abstract

McCabe and Halstead method-level metrics are among the well-known and widely used quantitative software metrics are used to measure software quality in a concrete way. Software defect prediction can guess which or which of the sub-modules in the software to be developed may be more prone to defect Thus, loss of labor and time can be avoided. The datasets which are used for software defect prediction, usually have an unbalanced class distribution, since the number of records with defective class can be fewer than the number of records with not defective class and this situation adversely affect the results of the machine learning methods. Information gain is employed in decision trees and decision tree based rule classifier and attribute selection methods. In this study, software metrics that provide important information for software defect prediction have been investigated and CM1, JM1, KC1 and PC1 datasets of NASA's PROMISE software repository have been balanced with the synthetic data over-sampling Smote algorithm and improved in terms of information gain. As a result, the software defect prediction datasets with higher classification success performance and the software metrics with increased information gain ratio are obtained in the decision trees.