集成学习融合依赖度的软件缺陷数量预测方法

Software Defect Quantity Prediction of Ensemble Learning Fusion Dependency

软件缺陷可能会导致软件产品故障和经济损失,有效识别潜在的软件缺陷在软件开发、运行维护过程中是至关重要的。先前软件缺陷预测方面的研究工作主要关注判断一个软件模块是否有缺陷。但由于测试资源,预测软件模块缺陷数会对软件开发和维护更有帮助。因此,在集成算法基础上,融合依赖度,提出一种新的缺陷数预测方法ELDDP(Ensemble Learning Fusion Dependencies for Software Defect Count Prediction)。首先,基于smote算法,对从最k近邻选择目标类方法进行改进。将类之间的依赖关系进行度量并引入smote算法中,选择目标类时优先选择依赖度高的类。其次,提出结合集成学习Adaboost.R2算法和Stacking模型融合算法构建缺陷数预测模型。在Promise数据集进行实验,对比了三种常用缺陷预测回归模型,实验结果表明该算法有较好的准确性和稳定性。

Software defects may lead to software product failures and economic losses,and effective identification of potential software defects is critical in the software development,operation and maintenance process.Previous research work in software defect prediction has focused on determining whether a software module is defective or not.However,due to testing resources,predicting the number of software module defects would be more helpful for software development and maintenance.Therefore,a new defect count prediction method ELDDP(Ensemble Learning Fusion Dependencies for Software Defect Count Prediction)is proposed based on the ensemble learning and fusion dependency.Firstly,the target class selection method from the k-most nearest neighbor is improved based on the smote algorithm.The dependency relationship between classes is measured and introduced into the smote algorithm,and the target classes with high dependency are selected in preference.Secondly,a combination of ensemble learning Adaboost.R2 algorithm and Stacking model fusion algorithm is proposed to construct a defect number prediction model.The experiment is carried on Promise dataset,and three commonly used defect prediction regression models are compared.It is showed that the proposed algorithm has better accuracy and stability.