Cross-Project Aging-Related Bug Prediction Based on Feature Transfer and Class Imbalance Learning

Software aging refers to system performance degradation and failure due to Aging-Related Bugs (ARBs) in long-running software systems. ARB prediction helps identify ARBs to prevent software aging. However, early project stages often lack ARB data for model training. To address this, cross-project ARB prediction (CPAP) is proposed, where a model is trained using labeled source projects to predict the ARB present in the target project. In CPAP, distribution discrepancy and class imbalance pose challenges, impacting model performance. In this paper, a hybrid CPAP combines feature transfer learning with class imbalance learning is proposed to solve the above problems, which we named KDK. In the feature transfer learning stage, KDK leverages Kernel Principal Component Analysis (KPCA) in conjunction with Double Marginalized Denoising Autoencoders (DMDA) to acquire feature representations that reduce distribution disparities among projects. In the class imbalance learning stage, K-means Clustering Cleaning Ensemble (KCE) is employed to rebalance and clean the source project dataset, who has a severe class imbalance and class overlap. Experiments are conducted on three open-source projects with two performance metrics (AUC and Balance). Experiment results demonstrate that our approach enhances Balance values by 9.0%, 32.9%, 4.4%, and 3.9% in comparison to state-of-the-art CPAP methods, namely TLAP, JPKS, SRLA, and JDA-ISDA, respectively.