A Novel Feature Selection Method Based on Maximum Likelihood Logistic Regression for Imbalanced Learning in Software Defect Prediction

The most frequently used machine learning feature ranking approaches failed to present optimal feature subset for accurate prediction of defective software modules in out-of-sample data. Machine learning Feature Selection (FS) algorithms such as Chi-Square (CS), Information Gain (IG), Gain Ratio (GR), RelieF (RF) and Symmetric Uncertainty (SU) perform relatively poor at prediction, even after balancing class distribution in the training data. In this study, we propose a novel FS method based on the Maximum Likelihood Logistic Regression (MLLR). We apply this method on six software defect datasets in their sampled and unsampled forms to select useful features for classification in the context of Software Defect Prediction (SDP). The Support Vector Machine (SVM) and Random Forest (RaF) classifiers are applied on the FS subsets that are based on sampled and unsampled datasets. The performance of the models captured using Area Ander Receiver Operating Characteristics Curve (AUC) metrics are compared for all FS methods considered. The Analysis Of Variance (ANOVA) F-test results validate the superiority of the proposed method over all the FS techniques, both in sampled and unsampled data. The results confirm that the MLLR can be useful in selecting optimal feature subset for more accurate prediction of defective modules in software development process.