基于旋转森林的分类器集成算法研究 - 计算机工程与应用.pdf

Computer Engineering and Applications 计算机工程与应用 2015 ，51（23 ） 149 基于旋转森林的分类器集成算法研究 邵良杉，马 寒 SHAO Liangshan, MA Han 辽宁工程技术大学 系统工程研究所，辽宁 葫芦岛 125105 System Engineering Institute, Liaoning Technical University, Huludao, Liaoning 125105, China SHAO Liangshan, MA Han. Study on classifier ensemble algorithm based on rotation forest. Computer Engineering and Applications, 2015, 51 （23 ）：149-154. Abstract ：In order to improve accuracy of decision tree, the rotation forest classifier ensemble algorithm based on feature transformation is introduced. To enhance the diversity of base classifiers and to improve predicting accuracy, the feature set is split into some subsets and principal component analysis is applied to each subset to create new subsample data for each base classifier. Using Weka, rotation forest is compared with Bagging and AdaBoost. Pruned and unpruned J48 decision tree are applied to be base classifier respectively and the average accuracy of 10 times 10 fold cross-validation is to be com- parison basis. The result shows that the accuracy of rotation forest is higher than others, and it proves that rotation forest is an effective classifier ensemble algorithm of decision tree. Key words ：rotation forest; classifier ensemble; principal components analysis; decision tree 摘 要：为提高决策树的集成分类精度，介绍了一种基于特征变换的旋转森林分类器集成算法，通过对数据属性集 的随机分割，并在属性子集上对抽取的子样本数据进行主成分分析，以构造新的样本数据，达到增大基分类器差异 性及提高预测准确率的目的。在Weka 平台下，分别采用Bagging 、AdaBoost 及旋转森林算法对剪枝与未剪枝的J48 决策树分类算法进行集成的对比试验，以10 次10折交叉验证的平均准确率为比较依据。结果表明旋转森林算法的 预测精度优于其他两个算法，验证了旋转森林是一种有效的决策树分类器集成算法。 关键词：旋转森林；分类器集成；主成分分析；决策树 文献标志码：A 中图分类号：TP 18 doi ：10.3778/j.issn. 1002-8331.1311-0252 1990 年，Schapire[1] 在PAC （Probably Approximately 值为最后输出分类结果。 [2] Correct）学习模型 的基础上分析并证明了弱学习算法与 基分类器的差异性是影响集成效果的一个重要因 [6] 强学习算法的等价性，即可通过集成，把好于随机猜测