A Comparative Study of Wrapper Feature Selection Techniques in Software Fault Prediction

Abstract

Software fault prediction aims to classify whether the module is defective or not-defective. In software systems, there are some software metrics may contain irrelevant or redundant information that leads to negative impact on the performance of the fault prediction model. Therefore, feature selection is an method that several studies have addressed to reduce computation time, improve prediction performance and provide a better understanding of data in machine learning. Additionally, the presence of imbalanced classes is one of the most challenge in software fault prediction. In this study, we examined the effectiveness of six different wrapper feature selection including Genetic Algorithm, Particle Swarm Optimization, Whale Optimization Algorithm, Cuckoo Search, Mayfly Algorithm and Binary Bat Algorithm for selecting the optimal subset of features. Then, we applied VanilaGAN to train the dataset with optimal features for handling the imbalanced problem. Subsequently, these generated training dataset and the testing dataset are fed to the machine learning techniques. Experimental validation has been done on five dataset collected from Promise repository and Precision, Recall, F1-score, and AUC are evaluation performance measurements.