Comparing Two Approaches for Adding Feature Ranking to Sampled Ensemble Learning for Software Quality Estimation

High dimensionality and class imbalance are two main problems that affect the quality of training datasets in software defect prediction, resulting in inefficient classification models. Feature selection and data sampling are often used to overcome these problems. Feature selection is a process of choosing the most important attributes from the original data set. Data sampling alters the data set to change its balance level. Another technique, called boosting (building multiple models, with each model tuned to work better on instances misclassified by previous models), is found also effective for resolving the class imbalance problem. In particular, RUSBoost, which integrates random undersampling with AdaBoost, has been shown to improve classification performance for imbalanced training data sets. In this study, we investigated an approach for combining feature selection with this ensemble learning (boosting) process. We focused on two different scenarios: feature selection performed prior to the boosting process and feature selection performed inside the boosting process. Ten base feature ranking techniques and an ensemble ranker based on the ten were examined and compared over the two scenarios. The experimental results demonstrate that the ensemble feature ranking method generally had better or similar performance than the average of the base ranking techniques, and more importantly, the ensemble method exhibited better robustness than any other base ranking technique. As for the two scenarios, the results show that applying feature selection inside boosting performed better than using feature selection prior to boosting.