An Approach for Improving Classification Accuracy Using Discretized Software Defect Data

Abstract

Predicting software defects in software systems at early stages of its development has always been a very crucial and desirable aspect of software development industry. Today the machine learning algorithms are playing a massive role in classifying and predicting the possible bugs during the design phase. In this research work, the authors have proposed a discretization method based on metrics threshold values in order to gain better classification accuracy on a given data set. For the experimentation purpose, the authors have chosen the defect data sets from NASA repositories. In this Jedit, Lucene, Tomcat, Velocity, Xalan, Xerces software systems have been considered for experimentation using WEKA. The authors have also considered object-oriented CK metrics specifically for the study. Two very common and popular classifiers namely Naive Bayes and voted perceptron for the classification purpose. In the proposed work, various performance measures like ROC, RMSE values have been considered and analyzed. The results show that classification accuracy improvements can be made while using proposed discretization method with both classifiers.