Two Stage Model to Detect and Rank Software Defects on Imbalanced and Scarcity Data Sets

摘要

In the software quality assurance process, it is crucial to prevent defective software to be delivered to customers since it can save the maintenance cost and increase software quality and reliability. Software defect prediction is recognized as an important process to automatically detect the possibility of having an error in the software. After defects are detected, it is then needed to identify their severity levels to avoid any effects that may obstruct the whole system. There were many trials attempts to capture errors by employing traditional supervised learning techniques. However, all of them are often faced with an imbalanced issue and scarcity of data, which causes decreased prediction performance. In this paper, we present a Two-Stage Model to detect and rank defects in software. The model focuses on two tasks. First, we will capture defects by applying an unbiased SVM called "R-SVM," which reduces a bias of the majority class by using the concept of threshold adjustment. Second, the detected modules will be ranked according to their severity levels by using our algorithm called "OS-YATSI," that combines semi-supervised learning and oversampling strategy to tackle the imbalanced issue. The experiment was conducted on 15 Java programs. The result showed that the proposed model outperformed all of the traditional approaches. In the defect prediction model, R-SVM significantly outperformed others on 6 programs in terms of F1. In the defect ranking model, OS-YATSI significantly outperformed all baseline classifiers on all programs at an average of 23.75% improvement in term of macro F1.