Hybrid Approach to Software Defect Prediction and Classification using a Variational Autoencoder and Scalable Graph Attention Network

Abstract

Software performance, security, and dependability depend on software defect prediction (SDP) and classification. Class imbalance, a lack of high-quality labeled data, overfitting in small or skewed datasets, poor interpretability, and the dynamic nature of software systems are among the issues that SDP classification faces. This research proposes an SDP with a classification framework that combines the scalable graph relearn attention network (SGRAN) and the optimized variational autoencoder (VAE) to tackle these challenges. Learning-to-rank under-sampling (LTRUS) selectively eliminates less-relevant majority-class instances of the majority class to reduce class imbalance. After learning structured latent representations, the VAE model improves generalization by identifying significant defect patterns. To further normalize the dataset by reducing undersampling effects and guaranteeing balanced representation between defective and non-defective modules, particle swarm optimization (PSO) is combined with crayfish bobcat optimization algorithm (CBOA) to dynamically optimize the termination condition. The expected defective modules are fed into the SGRAN for classification, ensuring unique and consistent defect categorization. The VAE-CBOA achieves 99.07% prediction accuracy and a high F1-score of 99.01%, outperforming previous studies. Using traditional methods, the proposed work achieved 98.2% recall and 98.02% accuracy. In each prediction and class responsibility, the proposed model’s efficiency and values are highlighted.