Predicting Defect-Prone Software Modules Using Generalized Discriminant Analysis-Based Feature Reduction Technique

Software Fault Prediction (SFP) strengthens the software development process by effectively dealing with the faults in software modules at the initial stages of software development, leading to lower maintenance costs besides achieving quality fault-free software. Unfortunately, SFP faces a few threats, such as data imbalance, high dimensionality, etc. To build an efficient fault prediction model, addressing the issue of high dimensionality in the software datasets is essential. Feature reduction (FR) has been an effective mechanism for dealing with the high dimensionality problem. This study suggests a novel FR procedure based on Generalized Discriminant Analysis (FRGDA) to single out the relevant traits only by reducing the dataset. The recommended FRGDA procedure was employed on four widely used supervised learning classifiers, namely Decision Tree (DT), K-nearest neighbour (KNN), Naive Bayes (NB), and Support Vector Machine (SVM) using six Java-based open-source static measure datasets from the PROMISE repository. This study used accuracy and mean error as the evaluation metrics to gauge the behaviour of the recommended procedure. Furthermore, the conduct of the recommended FRGDA procedure was contrasted with the extant FR techniques like Principal Component Analysis (FRPCA), Linear Discriminant Analysis (FRLDA), Kernel-based PCA (FRKPCA), and Quadrative Discriminant Analysis (FRQDA). The investigatory data demonstrated that the recommended FRGDA procedure bettered the model's predictive performance and outclassed other studied FR techniques.