Software Aging Defect Prediction Using Novel Oversampling Technique in Cloud Oriented Softwares

Software Aging refers to the performance degradation and increased failure rate in long running software systems and the bugs that cause this phenomenon are called Aging-Related Bugs (ARB). The main cause of this occurrence is the accumulation of run-time errors in the system due to memory leakage, null pointer exception, deadlock, etc. It is required to predict these ARBs using classification models. A high-performance classification model needs a lot of training data to be trained for ARB prediction. However, software aging is a class imbalance problem where non-ARB instances in dataset outnumber ARB instances. Oversampling techniques create duplicate instances result in low probability of accuracy or low probability of finding the defects i.e., less Recall value. In this study, novel Complexity-based Oversampling Technique (COSTE) is applied for ARB prediction in the cloud computing datasets where the complexity of each instance is measured to predict ARBs, which can attain high accuracy and recall value simultaneously in the prediction models. COSTE also performs better in terms of F1-score and precision in predicting ARB. The studies are carried out to compare COSTE results with those of SMOTE, another oversampling method. Feature selection techniques are used to select relevant features from cloud computing datasets in order to increase the efficacy of prediction models. Differential Evolution (DE), an optimization technique, is utilized to calculate the appropriate weight of each measure. Thus, as per our knowledge for the first time, COSTE is proved to be better oversampling technique as compared to SMOTE for predicting ARBs in cloud computing dataset. The Wilcoxon Signed Rank Test, a statistical test, shows that COSTE’s capacity to outperform SMOTE is significant.