摘要
支持向量回归机(SVR)在实际的学习应用中,由于数据时空的复杂性和算法本身的参数选择,学习模型难以达到预期的效果.针对这个问题,提出了基于Boosting集成的支持向量回归机方法.通过在原始数据集加权采样的基础上,进行多次迭代子SVR机器学习,不断调整样本权值再采样,优化机器学习模型,然后对迭代所得的每级支持向量回归结果按某种组合方法进行集成,得到最终的回归函数形式.应用该方法进行了仿真试验和滑坡变形时序预测研究.结果表明:使用集成的SVR进行回归预测较之单一的SVR具有更高的准确性和更好的泛化性.对Boosting与Bagging 2种不同的集成SVR,进行了比较研究,试验结果表明,2种算法性能相差不大,总体上前者强于后者.
The regression results of the practically implemented support vector machines （SVR） are often far from the theoretically expected level because of the high complexity of time and space of real data, and it is difficult to practically select hyper-parameters for SVR. To improve the limited regression performance of the real SVR, we proposed a method to form committee machines for regression by using boosting techniques. In the process of algorithm, each sub-SVR was trained independently and the training set was created based on weighted random sampling from the original dataset. The final SVR solution was produced by aggregating the sub-SVRs results through methods such as least-squares estimation based weighting etc. Experiments on both artificial and real-world datasets indicated that the Boosting ensemble SVR outperformed single SVR. In addition, we compared ensembles constructed using boosting with those using bagging, and the results showed that boosting was generally much better.
出处
《湖南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2007年第9期6-10,共5页
Journal of Hunan University:Natural Sciences
基金
贵州省交通建设科技项目(2003-318-80-201)
交通部西部重点基金资助项目(200331880201)
作者简介
董辉（1976-），男，湖南安乡人，中南大学博士研究生.通讯联系人，E-mail：aneurinsky@163．com．
