基于反向标定合成数据的改进集成算法.pdfVIP

第 33 卷第 6 期 电 子 与 信 息 学 报 Vol.33No.6 2011 年 6 月 Journal of Electronics Information Technology Jun. 2011 基于反向标定合成数据的改进集成算法 韩 敏* 朱新荣 (大连理工大学电子信息与电气工程学部 大连 116024) 摘 要：面对获得的数据量越来越多，需要处理的数据类型也不尽相同，因此就需要寻找一种具有较好泛化性能和 较高分类精度的算法。该文提出一种通过借用反向扩充训练数据样本对输入数据类型的不敏感性和径向基函数网络 模型快速学习的能力来进行集成的混合算法。采用渐进P值作为受试者特征曲线下面积与0.5 判断冗余特征的标准， 将反向标定合成的新数据对分类器进行训练，通过比较训练误差的变化来决定新分类器的添加，最终以绝大多数投 票方法对所有的分类器进行决策融合。最后以 UCI 数据为实验，结果表明该算法可以较好地适应于不同数据类型， 得到比其它集成算法更高的分类精度。 关键词：集成算法；径向基函数神经网络；反向扩充训练数据样本；投票法；ROC 曲线 中图分类号：TP274 文献标识码： A 文章编号：1009-5896(2011)06-1475-06 DOI: 10.3724/SP.J.1146.2010.00954 A New Ensemble Algorithm Based on Oppositional Relabeling of Artificial Data Han Min Zhu Xin-rong (Faculty of Electronic Information and Electrical Engineering, Dalian University of Technology, Dalian 116024, China) Abstract: The amount of data increases rapidly, and the types of data need to be handled become more and more various, a new algorithm with better generalization performance and higher classification accuracy is indispensable. In this paper, a new hybrid algorithm is proposed, which takes the advantage of the insensitivity to the input data of the Diversity Ensemble Creation by Oppositional Relabeling of Artificial Training Examples (DECORATE) algorithm and the efficiency of the radial basis functions neural network model. Asymptotic P value to decide the relationship between the area under receiver operator characteristic with 0.5 which belong to redundant features, and the oppositional relabeling artificial data is used to train the classifier. Then the new classifier is added which will lower