有限标注数据下的软件故障倾向预测方法  

Software Fault-proneness Prediction with Limited Labeled Data

作　　者：罗云锋[1] 贲可荣 

机构地区：[1]海军工程大学计算机工程系,武汉430033

出　　处：《武汉理工大学学报》2010年第20期178-183,共6页Journal of Wuhan University of Technology

基　　金：国家预研项目(513270104)

摘　　要：将直推式支持向量机用于预测有限标注样本下的模块故障倾向。针对软件故障样本中存在正负样本数量不平衡、比例不确定等特点,对训练样本的惩罚项引入权重因子,并采用单样本调整准则以解决直推式支持向量机没有考虑样本不平衡、需要事先确定未标注样本正负比例等问题。在7种不同规模的公开失效数据集上的实验结果表明:改进的直推式支持向量机算法利用了未标注数据隐含的信息,能够获得比归纳式分类算法更好的性能。The transductive support vector machine（TSVM） is used to predict the fault-proneness of the software modules with limited labeled data.Given that the software fault datasets are imbalanced,and the ratio of positive samples to negative samples is uncertainly,an improved TSVM is constructed.In our model,the penalty items for the training samples are weighted,and an individually changing criterion is utilized to solve the problem that the number of positive testing samples must be appointed before training in TSVM.Experimental results,based on seven public data set of software faults with different sizes of modules,indicate that compared to the inductive classification models with the available set of labeled program modules,the improved TSVM exploits the knowledge stored in the software metrics of the unlabeled software modules,and improves performance of software fault-proneness prediction.

关 键 词：软件故障倾向预测 软件度量 直推式支持向量机 有限标注数据 

分 类 号：TP306[自动化与计算机技术—计算机系统结构]