结合特征对齐与实例迁移的跨项目缺陷预测  

Cross-project defect prediction combining feature alignment and instance migration

作　　者：李莉[1] 赵鑫[1] 石可欣 苏仁嘉 任振康 Li Li;Zhao Xin;Shi Kexin;Su Renjia;Ren Zhenkang(College of Computer&Control Engineering,Northeast Forestry University,Harbin 150040,China)

机构地区：[1]东北林业大学计算机与控制工程学院,哈尔滨150040

出　　处：《计算机应用研究》2023年第10期3091-3099,共9页Application Research of Computers

基　　金：黑龙江省教育科学规划课题(GJB1421251)。

摘　　要：为解决跨项目缺陷预测中源项目和目标项目分布差异较大的问题,提出了一种基于特征对齐和实例迁移的两阶段缺陷预测方法(FAIT)。首先,在特征对齐阶段,根据边缘概率分布进行特征的边缘分布对齐;然后,基于源项目和目标项目构建条件分布映射矩阵完成条件分布对齐;最后,在实例迁移阶段,通过改进了权重调整策略的TrAdaBoost方法构建跨项目缺陷预测模型。以F 1作为评价指标,当目标项目有标签实例比例为20%时,FAIT性能最佳,且两过程特征对齐优于单一过程特征对齐。此外,FAIT的预测性能在AEEEM和NASA数据集上分别提高了10.69%、15.04%。FAIT在一定程度上解决了源项目与目标项目的分布差异,能够取得较好的缺陷预测性能。To address the problem of significant distributional differences between source and target projects in cross-project defect prediction,this paper proposed a two-stage defect prediction method based on feature alignment and instance transfer(FAIT).In the feature alignment stage,FAIT aligned the marginal distributions of features based on their probability distributions.Next,this method constructed a conditional distribution mapping matrix based on both source and target projects to achieve conditional distribution alignment.Finally,in the instance transfer stage,it built a cross-project defect prediction model using an improved weight adjustment strategy for TrAdaBoost.FAIT achieves the best performance when the proportion of labeled instances in the target project is 20%,with F 1 as the evaluation metric.Furthermore,dual-process feature alignment outperforms single-process feature alignment,and FAIT’s predictive performance improves by 10.69%and 15.04%on the AEEEM and NASA datasets,respectively.To some extent,FAIT addresses distributional differences between source and target projects and achieves good defect prediction performance.

关 键 词：跨项目缺陷预测 特征对齐 最大均值差异 实例迁移 TrAdaBoost 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]