不平衡数据的软件缺陷预测方法

数据的不平衡问题是数据分类领域中的一个热点问题。当分类算法处理这些数据时,算法将偏向多数类而忽视少数类。在软件缺陷预测领域,它并没有引起足够的重视,在4组NASA不平衡数据上比较和分析了14种分类算法的性能,为了克服数据的不平衡性,采用SMOTE技术对软件缺陷数据进行平衡化,最后在AUC和F-measure评价指标下对算法进行评估。实验结果表明随机森林算法在4组数据上表现最佳,这为软件缺陷预测提供了很好的参考。

Research on Software Defect Prediction with Imbalanced Data

The problem of imbalanced data is well known in many applications of classification learning.But it doesn’t have significant attention in the filed of software defect prediction.Software defect data usually have data imbalance problems due to the fact that one class is represented by a much larger number of instances than other classes.Consequently,algorithms are biased to wards majority class and away from minority class.This paper investigates the capability of fourteen different classification algorithms,compares four software defect imbalanced data from NASA Metrics Data Programs.In terms of AUC and F-measure,the results show that Random Forest algorithm is the best performer after balancing.This result provides reference for software quality prediction.