摘要
针对软件可靠性早期预测中软件复杂性度量属性维数灾难问题,提出了一种基于最小绝对值压缩与选择方法(The Least Absolute Shrinkage and Select Operator,LASSO)和最小角回归(Least Angle Regression,LARS)算法的软件复杂性度量属性特征选择方法。该方法筛选掉一些对早期预测结果影响较小的软件复杂性度量属性,得到与早期预测关系最为密切的关键属性子集。首先分析了LASSO回归方法的特点及其在特征选择中的应用,然后对LARS算法进行了修正,使其可以解决LASSO方法所涉及的问题,得到相关的复杂性度量属性子集。最后结合学习向量量化(Learning Vector Quantization,LVQ)神经网络进行软件可靠性早期预测,并基于十折交叉方法进行实验。通过与传统特征选择方法相比较,证明所提方法可以显著提高软件可靠性早期预测精度。
To cope with the software complexity metric attributes dimension disaster which exists in the software relia- bility early prediction, this paper put forward a software complexity metric attribute feature selection method based on Least Absolute Shrinkage and Selection Operator（LAS~））method and the Least Angle Regression（LARS）algorithm. This method can filter out some software complexity metric attributes which have smaller influence on the early predic- tion results and can obtain the key attributes subsets associated most closely with the prediction result. This paper firstly analyzed the characteristics of LAS~） regression method and its application in feature selection, secondly modified the LARS algorithm so that it can be used to solve the problems which LASSO method involves and get relevant complexity metric attribute subsets, lastly combined with the Learning Vector Quantization（LVQ）neural network to carry on the early software reliability prediction experiment. During the experiment, the authors used the 10-fold experiment meth- ods. The experiment results indicate that the method can improve early prediction accuracy of software reliability.
出处
《计算机科学》
CSCD
北大核心
2013年第11期169-173,共5页
Computer Science
基金
国家863项目计划(2008AA01Z404)资助
