A Software Aging-Related Bug Prediction Framework Based on Deep Learning and Weakly Supervised Oversampling

Abstract

Software aging refers to the phenomenon of sys-tem performance degradation and eventual failure caused by Aging-Related Bugs (ARBs). Software aging seriously affects the reliability and availability of software systems. To discover and remove ARBs, ARBs prediction is presented, and most of them only employed static code metrics to predict those buggy codes. However, static code metrics do not capture the syntactic and semantic features of the code, which are important to building accurate prediction models. To address this problem, we design a deep neural network by combining the bidirectional long short-term memory (BLSTM) and the attention mechanism to extract context-sensitive semantic features of the code. In addition, we apply a weakly supervised oversampling (WSO) method to alleviate class imbalance problems in datasets. We named our framework ABLSTM-WSO. We conduct experiments with five classifiers on two widely used open-source projects(MySQL and Linux) and use AUC, Balance, and F1-score as the evaluation metrics. Experimental results show that ABLSTM-WSO can significantly improve the ARBs prediction performance.