软件缺陷预测中基于Wrapper的特征选择方法

Feature Selection Based on Wrapper During Software Defect Prediction

特征选择是提高软件缺陷预测精度的关键步骤之一。传统的软件缺陷预测过程主要基于Filter方式进行特征选择,基于Wrapper特征选择方法的研究还处于起步阶段。为了进一步研究Wrapper式特征选择方法在软件缺陷预测中的应用情况,将特征选择和缺陷预测过程相融合,结合不同的评价指标,设计了8种基于Wrapper式特征选择的缺陷预测方法。在这些方法中,首先选择4种常用的缺陷预测算法分别作为内部与外部分类器,然后在AUC和F-measure指标下选择特征子集,在AUC指标下评估预测结果。仿真结果表明,内部分类器和外部分类器均选择为RF时,软件缺陷预测精度最佳,NB次之,但是RF耗费时间较多,综合考虑精度与效率,推荐内外分类器均采用NB算法。

Feature selection is one of the key steps to improve prediction accuracy of softwaredefects. In the traditional fields of software defect prediction,feature selection based on filter is moreoften used than wrapper. In order to further investigate how to use feature selection based wrapper insoftware defect prediction application,eight feature selection methods based on wrapper with combiningthe process of prediction and selection is designed. Firstly,four common defect prediction algorithms areused as internal and external classifiers. Then AUC and F-measure are used to assess the performance offeature selection methods. Simulation results show that the internal and external classifiers select RF,software defect prediction accuracy is the best,and NB is the second. But because of RF is more time-consuming,NB algorithm is recommended considering the precision and efficiency.