Deep Learning Architectures for Software Fault Prediction: The Impact of Error‐Type Metrics and Class Imbalance

ABSTRACT Software fault prediction (SFP) plays a crucial role in modern software development by enabling early identification of fault‐prone modules and efficient allocation of testing resources. While deep learning approaches have shown promise in this domain, challenges persist regarding architectural choices, metric selection, and class imbalance issues. This study presents a comprehensive comparison between deep neural networks (DNNs) and hybrid Graph Neural Network‐Long Short‐Term Memory (GNN+LSTM) models for SFP, investigating their effectiveness when combined with both conventional software metrics and error‐type metrics. We evaluate these approaches on four real‐world Java projects: ANTLR v4, JUnit, OrientDB, and Elastic Search. Our results demonstrate that GNN+LSTM models consistently outperform traditional DNN approaches, achieving improvements of up to 4% in accuracy and 4% in F1‐score. However, we identify challenges in combining different metric sets, with performance actually degrading compared to our previous study using error‐type metrics alone, suggesting potential multi‐collinearity issues. Additionally, we examine the effectiveness of the synthetic minority oversampling technique (SMOTE) in addressing the class imbalance issue, observing improvements of up to 6.6% in accuracy for GNN+LSTM models in severely imbalanced datasets. Our findings provide practical insights for selecting appropriate model architectures and metric combinations in SFP while highlighting the importance of carefully considering feature interactions and class imbalance mitigation strategies.