Handling Imbalanced Data using Ensemble Learning in Software Defect Prediction

摘要

With the ever growing software industry, software defect prediction is one of the key ingredients in recipe of producing good quality software. Defects uncovered well in time helps in saving resources in terms of time, effort and money. However imbalanced nature of software data may hamper the resultant performance of models leading to incorrect interpretations of results. This problem has dragged attention of researchers and many solutions are proposed to overcome the effect of this problem. This paper aims to provide empirical comparison of software defect prediction models developed by using various boosting based ensemble methods on three open source JAVA projects. Four ensemble methods incorporate resampling techniques within them. Performances of models obtained are evaluated using stable metrics like Balance, G-Mean and AUC. Results show that use of resampling techniques before classifying using ensemble method has significantly improved model prediction as compared to classic boosting models. RUSBoost is the undisputed winner amongst all followed by MSMOTEBoost and SMOTEBoost.