Within-Project Software Aging Defect Prediction Based on Active Learning

Abstract

Long-running software systems tend to exhibit performance degradation and increase failure rate, and the phenomenon is known as software aging. The bugs that cause the aging phenomenon are called Aging-Related Bugs (ARBs), and may bring serious economic loss or even endanger human security. To discover and remove ARBs, ARBs prediction is presented. But ARBs prediction model often needs a large number of training data in order to train a high performance classification model. In practice, the labeled data are rare in many cases. In addition, it is difficult to label all samples manually. Furthermore, there is a serious class imbalance problem in ARBs datasets. In order to address the two problems, we propose a framework named QUIRE-HUE. On the one hand, we use a approach named Active Learning by Querying Informative and Representative Examples (QUIRE) to select a few informative and representative samples to label for training set, which can reduce the cost of labeling and get a high performance classification model. On the other hand, we apply a Hashing-Based Undersampling Ensemble (HUE) by constructing diversified training subspaces for undersampling to alleviate class imbalance problem. A set of experiments are performed on two large open-source projects (MySQL, Linux) with six different machine learning classifiers. We use Balance and AUC as the evaluation metrics. Experimental results indicate that QUIRE-HUE achieves encouraging results. Average AUC and Balance are 0.769 and 0.812 respectively on MySQL dataset, 0.772 and 0.828 respectively on Linux dataset, which significantly outperforms all baseline methods.