Handling High-Dimensionality on Software Defect Prediction with FLDA

For years, Software Defect Prediction (SDP) has been used as a way to improve software reliability. It is used as a tool to detect the defects on software module before the testing phase. The steps consist of building machine learning model by training dataset with classifier and predict on defective modules. Many datasets suffer from high dimensionality because of the high number of features and those features are mostly irrelevant to predict defect. This results the data to be unnecessarily bulky and the classification process to be time-consuming. We propose a feature extraction technique called FLDA for handling dimensionality problem of dataset and improving the classification performance. We use a total of four dataset from NASA MDP. For classifiers, we use Support Vector Machine (SVM), Random Forest (RF), Naive Bayes (NF), and Multi Layer Perceptron (MLP). Based on the study results, FLDA can significantly reduced the dimension of datasets by creating new feature that contains the most relevant information. FLDA can also shorten the processing time of the classifiers. When compared to another feature extraction technique such as PCA, FLDA can easily outperform it in terms of Recall and AUC.