面向代码演化的集成软件缺陷预测模型

Integrated Software Defect Prediction Model for Code Evolution

不同版本的软件缺陷之间存在一定的关联性,在面向演化项目的缺陷预测方面通常通过构建预测模型,以历史版本缺陷数据为输入,对后续版本的缺陷进行预测,但普遍存在缺陷预测性能较差的问题。针对该问题,提出了一种面向代码演化的集成软件缺陷预测模型,通过选择与缺陷相关联的代码度量元以及版本间的演化度量元,由决策树(J48)、逻辑回归(LR)、神经网络(NN)、朴素贝叶斯(NB)各自迭代产生分类器,结合Adaboost集成学习方法,使其在训练分类器时更关注每一轮的错分元组,得到不同的预测集成模型。在PROMISE软件数据集上实验表明,针对代码演化问题,集成后的模型比单一的机器学习模型在精确度、召回率、F1和AUC上分别都提高了20.3%、44.9%、43.4%、45.3%,其中基于J48分类器的集成模型预测性能最好,比基于LR、NN、NB分类器的集成模型在AUC指标上平均提高7.7%、4.6%、8.0%,与近年来面向演化项目的缺陷预测技术对比,结果表明本文中集成方法更有效。

There is a certain correlation between software defects of different versions.In the aspect of defect prediction for evolu⁃tion projects,the defects of subsequent versions are predicted by building a prediction model and taking the defect data of historical versions as the input,but the defect prediction performance is generally poor.To solve this problem,this paper proposes an inte⁃grated software defect prediction model for code evolution.By selecting the code metric associated with the defect and the evolu⁃tion metric between versions,the classifier is iteratively generated by decision tree(J48),logistic regression(LR),neural network(NN)and Naive Bayes(NB),combined with AdaBoost integrated learning method.When training the classifier,it pays more at⁃tention to the misclassification tuples of each round,and obtains different prediction integration models.Experiments on promise software data set show that for the problem of code evolution,the integrated model improves the precision,recall,F1 and AUC by 20.3%,44.9%,43.4%and 45.3%respectively compared with the single machine learning model.Among them,the integrated model based on j48 classifier has the best predictability,which is 7.7%,4.6%and 8.0%higher than the integrated model based on LR,NN and NB classifier,compared with the defect prediction technology for evolution projects in recent years,the results show that the integration method in this paper is more effective.