软件缺陷倾向性预测投票方法

Software defect proneness prediction voting method

软件缺陷预测方法可以在项目的开发初期,通过预先识别出可能含有缺陷的软件模块来优化测试资源的分配。而其中一些研究通常是基于单分类器数据切分后性能的平均值来度量模型的,但是由于数据集存在不平衡等问题,其得到的性能并没有得到很大改善,而且在此过程中并不能得到被测样本最终的预测结果,即是否有缺陷。故基于此提出了一种软件缺陷倾向性预测投票方法,即首先对数据集进行主成分分析(PCA),并对数据进行m×2分层交叉验证,结合下采样技术进行建模,对于得到的所有预测结果进行投票,得到最终的预测结果,基于该结果计算模型的性能。该方法在NASA的KC1,PC3,PC4数据集上进行了实验,使用了两种分类器:logistic回归模型和决策树模型,并且对该方法得到的性能和平均的性能进行了对比,以precision、recall、F1值、accuracy作为性能指标。实验结果表明该方法所得到的性能有一定的提升。

Software defect prediction methods can optimize test resources allocation by identifying defects modules in the early stage of project development.Some studies usually measure the model based on the average performance of a classifier after data segmentation.However,due to the dataset's imbalance problem,its performance has not been greatly improved,and it cannot obtain final prediction result of test samples,that is,whether there is defective.Therefore,this paper proposed a software defect proneness prediction voting method,that is,adopted principal component analysis first,and processes m×2 hierarchical cross-validation,builded model combined with under-sampling technology,voted on all obtained prediction results to get the final prediction result,and calculated the performance of the model based on these results.This method is conducted on NASA's KC1,PC3,and PC4 data sets using two classifiers:logistic regression model and decision tree model.The performance measures are precision,Recall,F1 value,and accuracy in this paper,compared with the average performances.The experimental results show the method's performances have some improvement.