低关联度的Boosting类集成算法研究  

Research of boosting algorithm with low correlation

作　　者：关超[1] 高敬阳[1] 尚颖[1] 

机构地区：[1]北京化工大学信息科学与技术学院计算机系,北京100029

出　　处：《计算机工程与设计》2012年第3期1112-1116,共5页Computer Engineering and Design

基　　金：国家自然科学基金项目(61074153)

摘　　要：针对Boosting类算法生成的个体网络的迭代方式相关性较高,对某些不稳定学习算法的集成结果并不理想的情况,基于Local Boost算法局部误差调整样本权值的思想,提出了基于距离及其权值挑选邻居样本的方法,并通过局部误差产生训练样本种子,采用Lazy Bagging方法生成针对各样本种子的个体网络训练样本集来训练、生成新的个体网络,UCI数据集上实验结果表明,该算法得到的个体网络相关度较小,集成性能较为稳定。Boosting is some kind of algorithm which may perform worse for its high correlation. A new algorithm based on Local- Boost algorithm is provided. The new algorithm adopted a new strategy which pick some ＂difficult＂ samples called seeds from their neighborhood samples which is selected according to the nearest weighted distance to generate new distribution. The new hypothesis is gotten through training this new distribution by LazyBagging method. Experimental result shows that the new al- gorithm is better than origins either in correlation or stability.

关 键 词：低相关度 神经网络集成 邻域误差 二次集成 Boosting集成 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]