基于联合特征分布匹配的跨项目缺陷预测

Cross-project defect prediction based on joint feature distribution matching

为解决跨项目软件缺陷预测研究中存在的特征不完备和分类边界模糊问题,提出一种基于联合特征的双编码器分布匹配方法(DeDM-JF)。利用卷积神经网络提取代码中与缺陷有关的结构语义特征,将其与人为选取的Handcrafted特征结合,形成联合特征;在此基础上,构建包含分布差异匹配层的双自编码器,学习跨项目全局和局部可迁移特征用于训练缺陷预测模型。面向软件缺陷数据仓库中的798对跨项目缺陷预测任务开展实验,与相关的跨项目缺陷预测方法比较,DeDM-JF方法预测的F-measure和MCC指标有明显提升。

To solve the problems of feature incompleteness and classification boundary ambiguity in cross-project software defect prediction,a joint feature-based dual-encoder distribution matching method(DeDM-JF)was proposed.Convolutional neural networks were used to extract defect-related structural semantic features in codes,and they were combined with handcrafted features to form joint features.On this basis,two autoencoders including distribution matching layers were constructed to learn the global and local transferable feature across projects for prediction model training.Experiments on 798 pairs of cross-project defect prediction tasks were conducted in the software defect data warehouse.Compared with the related cross-project defect prediction methods,the F-measure and MCC predicted using DeDM-JF are significantly improved.