应用非线性加权的集成学习软件缺陷序列预测算法  

Prediction Algorithm for Software Defect Series Based on Nonlinear Weighted Ensemble Learning

作　　者：贾晓琳[1] 樊帅帅 罗雪[1] 朱晓燕[1] JIA Xiaolin;FAN Shuaishuai;LUO Xue;ZHU Xiaoyan(School of Electronics and Information Engineering, Xi＇an Jiaotong University, Xi＇an 710049, China)

机构地区：[1]西安交通大学电子与信息工程学院学院,西安710049

出　　处：《西安交通大学学报》2017年第7期156-161,共6页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(61402355);中央高校基本科研业务费专项基金资助项目(jj2014050)

摘　　要：针对当前软件缺陷序列预测算法准确度不高的问题,提出了基于非线性加权的集成学习软件缺陷序列预测算法(NLWEPrediction)。该算法在常见线性集成预测算法的基础上增加了非线性回归项,回归项代表了集成预测算法中基预测算法之间的相互关系,修正了线性集成预测的偏差,并通过梯度下降法求解了模型中的参数。实验表明:NLWEPrediction在14个软件缺陷数据集上的均方误差均小于250,并且平均绝对误差均小于13。通过与基预测算法、集成预测Bagging、Stacking算法和只考虑两个分类器关系的非线性加权集成学习算法进行对比,可以看出,NLWEPrediction预测算法的均方误差和平均绝对误差显著减小,预测精度显著提高,说明在线性集成预测算法基础上增加非线性回归项,能够有效提高集成学习算法的分类效果。Aiming at the problem of the basic prediction algorithm with relative low prediction accuracy, a novel and improved prediction algorithm NLWEPrediction is proposed based on non- linear weighted and ensemble learning. It combines the advantages of linear ensemble learning and the relationship between the base predict algorithms, which corrects the prediction deviation and uses gradient descent method to calculate the model parameters. The experiments proved that the NLWEPrediction＇s mean squared error in datasets is lower than 250, and the mean absolute difference is lower than 13. The algorithm was compared with its four base prediction algorithms, other two ensemble prediction algorithms Bagging and Stacking and original NLWEPrediction for efficiency analysis. Experimental results showed that NLWEPrediction has obviously low mean square error and average absolute error. The prediction accuracy is improved. So, adding the nonlinear regression terms can improve the capability of ensemble clas- sifier.

关 键 词：软件缺陷序列 预测算法 软件缺陷 集成学习 

分 类 号：O121.8[理学—数学]