一种半监督集成跨项目软件缺陷预测方法

发布时间：2018-05-20 11:31

本文选题：跨项目软件缺陷预测 + 半监督学习　；参考：《软件学报》2017年06期

【摘要】：软件缺陷预测方法可以在项目的开发初期,通过预先识别出所有可能含有缺陷的软件模块来优化测试资源的分配.早期的缺陷预测研究大多集中于同项目缺陷预测,但同项目缺陷预测需要充足的历史数据,而在实际应用中,可能需要预测项目的历史数据较为稀缺,或这个项目是一个全新项目.因此,跨项目缺陷预测问题成为当前软件缺陷预测领域内的一个研究热点,其研究挑战在于源项目与目标项目数据集间存在的分布差异性以及数据集内存在的类不平衡问题.受到基于搜索的软件工程思想的启发,提出了一种基于搜索的半监督集成跨项目软件缺陷预测方法 S~3EL.该方法首先通过调整训练集中各类数据的分布比例,构建出多个朴素贝叶斯基分类器;随后,利用具有全局搜索能力的遗传算法,基于少量已标记目标实例对上述基分类器进行集成,并构建出最终的缺陷预测模型.在Promise数据集及AEEEM数据集上与多个经典的跨项目缺陷预测方法(Burak过滤法、Peters过滤法、TCA+、CODEP及HYDRA)进行了对比.以F1值作为评测指标,结果表明:在大部分情况下,S~3EL方法可以取得最好的预测性能.
[Abstract]:The software defect prediction method can optimize the allocation of test resources by pre-identifying all software modules that may contain defects at the beginning of the project development. Most of the early research on defect prediction focuses on the same project defect prediction, but the same project defect prediction needs sufficient historical data, but in practical application, the historical data may be scarce. Or this project is a new project. Therefore, cross-project defect prediction has become a research hotspot in the field of software defect prediction. The research challenge lies in the difference of distribution between the source project and the target project data set and the class imbalance problem in the data set. Inspired by the idea of search-based software engineering, a search-based semi-supervised integrated cross-project software defect prediction method, Sz3ELL, is proposed. Firstly, by adjusting the distribution ratio of all kinds of data in the training set, a number of naive Bayesky classifiers are constructed, and then the genetic algorithm with global searching ability is used. Based on a small number of labeled target examples, the base classifier is integrated and the final defect prediction model is constructed. The Promise data sets and AEEEM data sets are compared with several classic cross project defect prediction methods: Burak filter and Peters filter. Using F1 value as the evaluation index, the results show that in most cases, the best prediction performance can be obtained by using the Szc3EL method.
【作者单位】：天津大学软件学院软件工程系;南通大学计算机科学与技术学院;
【基金】：国家自然科学基金(61202030,61373012,61202006,71502125)~~
【分类号】：TP18;TP311.53

【相似文献】