Software Defect Prediction Using Generative Adversarial Network: A Comparative Study of Semi-Supervised Machine Learning Models

Software defect prediction is a crucial aspect of software quality assurance that facilitates early detection and prevention of potential defects, minimizing the cost and adverse consequences associated with software failures. However, it always suffers from the class imbalance problem, which makes it difficult for the model to generalize from minority defect instances. As a result, models may become biased toward the majority class. In this paper, we propose a software defect prediction framework that integrates generative adversarial network-based data augmentation with semi-supervised machine learning (SSML) techniques. To predict software defects, both traditional machine learning and SSML algorithms are applied in this study. The implemented techniques include k-nearest neighbors, support vector machine, decision tree, random forest, label propagation, and artificial neural network, with performance evaluated using accuracy, precision, recall, and F1-score. A comparative analysis has been performed between conventional machine learning and SSML approaches based on six benchmark datasets: CM1, Linux-driver net, Eclipse 3.0, JM1, KC1, and Prop-1. The results show that random forest achieved the highest accuracy (96.67%) and F1-score (94.38%) on the Linux-driver net dataset. These findings confirm the potential of SSML models combined with tabular generative adversarial network-based augmentation in enhancing defect prediction performance.