Software Defect Prediction with Skewed Data

Software defect prediction is often employed as a cost-effective tool to focus quality-improvement resources on poor-quality program modules. Many software measurement and defect data sets are very skewed, where the proportion of not-fault-prone (majority class) modules is substantially larger than that of fault-prone (minority class) modules. Data sampling and Boosting (Boost) are useful techniques for alleviating this problem. While various Data Sampling techniques are available, our prior studies have shown Random Undersampling (RUS) to be effective. RUS works by randomly removing instances from the majority class of the training data. This paper investigates combining Data Sampling and Boosting (RUSBoost) for software defect prediction with skewed software measurement and defect data sets. We consider two variations of RUSBoost depending on the percentage of fault-prone modules desired in the post-sampling training data set. Labeled as RUSBoost_A and RUSBoost_B, they respectively represent whether 35% or 50% of the post-sampling training data set should be fault-prone modules. A case study of 15 data sets from multiple real-world projects is used to demonstrate that RUSBoost performs significantly better than a model built without any Data Sampling or Boosting technique. Moreover, RUSBoost_B significantly outperforms RUSBoost_A.