IMPLICATIONS OF DISCRETIZATION TOWARDS IMPROVING CLASSIFICATION ACCURACY FOR SOFTWARE DEFECT DATA

摘要

Since the advent of new software architectures, paradigms and technologies the software design and development has developed a cutting edge requirements of being on the right track in terms of software quality and reliability. This leads the prediction of defects in software at its early stages of its development. Implications of machine learning algorithms are now playing a very crucial role in classification and prediction of the possible bugs during the systems design phase. In this research work a discretization method is proposed based on the Object Oriented metrics threshold values in order to gain better classification accuracy on a given data set. For the experimentation purpose, Jedit, Lucene, tomcat, velocity, xalan and xerces software systems from NASA repositories have been considered and classification accuracies have been compared with the existing approaches with the help of open source WEKA tool. For this study, the Object Oriented CK metrics suite has been considered due to its wide applicability in software industry for software quality prediction. After experimentation it is found that Naive Bayes and Voted Perceptron, classifiers are performing well and provide highest accuracy level with the discretized dataset values. The performance of these classifiers are checked and analyzed on different performance measures like ROC, RMSE, Precision, Recall values in this research work. Result shows significant performance improvements towards classification accuracy if used with discrete features of the individual software systems.