A Novel Model Based on Nonlinear Manifold Detection for Software Defect Prediction

Abstract

Development of software that may be encouraging for the developers and yield more customer satisfaction in lesser time and cost requires early prediction of defects lying already in the software system. Development of a defect-free and reliable software system involves conducting series of test cases which is actually a time consuming and cost oriented exercise. It requires framing a defect prediction model applying effective technique with suitable defect prediction performance measures that may be empirically validated for ensuring relevance to software organizations. Although series of defect prediction models have been developed using various classifiers and different techniques on defect datasets but those models were not at all fault-free and fully effective to achieve the goal. As such, it has become pertinent to set up an empirical framework and develop a newer Nonlinear Manifold Detection (NMD) Model along with various machine learning approaches for prediction of defects in software in most accurate manner. The new NMD Model ventured in identifying the attributes which are best and in that process all the unwanted, redundant and undesired attributes were eliminated. In this model, critical analysis and comparison with other Feature selection approaches have been made and the results have showed that NMD Model is more accurate and effective in predicting software defects. The prediction performance of various machine learning approaches have actually been compared by using measures like Accuracy, MAE, RMSE, AUC and they have also been tested statistically by use of Friedman and Nemenyi test. The experiment finally proved that NMD Model is more effective, significant and better result-oriented in terms of accuracy than other defect prediction approaches.