一种变粒度缺陷报告严重程度预测方法  

A severity predication method for variable granularity bug report

作　　者：贾焱鑫 林浩 陈翔[1] 濮雪莲 葛骅 JIA Yanxin;LIN Hao;CHEN Xiang;PU Xuelian;GE Hua(School of Information Science and Technology,Nantong University,Nantong 226019,China;School of Economics and Management,Nantong University,Nantong 226019,China)

机构地区：[1]南通大学信息科学技术学院,江苏南通226019 [2]南通大学经济与管理学院,江苏南通226019

出　　处：《南通大学学报（自然科学版）》2022年第1期44-51,共8页Journal of Nantong University(Natural Science Edition)　

基　　金：国家自然科学基金青年科学基金项目(61202006)。

摘　　要：已有工作一般将缺陷报告严重程度预测问题建模为粗粒度分类问题(即二分类问题)或细粒度分类问题(即多分类问题),更细的预测粒度意味着对分类准确性提出更高的要求,因而预测性能通常会下降。通过有效利用粗粒度分类的高预测性能与细粒度分类更为精细的预测粒度,提出一种变粒度缺陷报告严重程度预测(variable granularity bug report severity prediction,VG-BSP)方法,VG-BSP方法首先对缺陷报告进行基于二分类的粗粒度划分,然后根据二分类的结果,选择对应的预测模型以预测出具体的严重程度(即基于多分类的细粒度划分)。为了验证VG-BSP方法的有效性,搜集了来自Eclipse的两个子项目JDT和CDT的缺陷报告和来自Mozilla项目的缺陷报告;并将VG-BSP方法与逻辑回归(logistic regression,LR)方法、K近邻(K-nearest neighbors,KNN)方法、决策树(decision tree,DT)方法、随机森林(random forest,RF)方法、支持向量机(support vector machine,SVM)方法等经典细粒度分类方法进行了对比。实验结果表明:基于macro-F1指标,VG-BSP方法的性能与已有的细粒度方法相比平均可提升19.6%。除此以外,使用LR分类算法和Spacy词向量可以使得VG-BSP方法取得最好的性能。The previous studies generally model the bug report severity prediction problem as a coarse-grained classification problem(i.e., binary classification problem) or a fine-grained classification problem(i.e., multi-class classification problem). Finer prediction granularity means higher requirements for classification accuracy, so prediction performance tends to decline. Since using the coarse granularity classification can achieve high prediction performance while using the fine granularity can obtain finer prediction granularity, a variable granularity bug report severity prediction method(VG-BSP) is porposed. Firstly, the bug report is classified into two categories via the coarse-grained classification. Then according to the previous classification results, the corresponding prediction model is utilized to predict the specific severity of the bug report via fine-grained multi-class classification. To verify the effectiveness of VG-BSP, bug reports from two Eclipse subprojects(i.e., JDT and CDT), and bug reports from the Mozilla project are collected. The proposed VG-BSP method is compared with the previous classical fine-grained classification methods,such as logistic regression(LR), K-nearest neighbor(KNN), decision tree(DT), random forest(RF), support vector machine(SVM). The results show that compared with the previous methods, the performance of the proposed method can be improved by 19.6% on average in terms of the macro-F1 measure. Moreover, the LR classifier and the Spacy word embedding can enable VG-BSP to achieve the best performance.

关 键 词：软件仓库挖掘 缺陷报告 严重程度预测 机器学习 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]