一种提高神经网络集成差异性的学习方法

集成学习已经成为机器学习的研究方向之一,它可以显著地提高分类器的泛化性能.本文分析了Bagging及AdaBoost集成方法,指出了这两种方法的缺陷;然后提出了一种新的基于神经网络的分类器集成方法DBNNE,该方法通过生成差异数据增加集成的差异性;另外,当生成一个分类器后,采用了测试方法确保分类器集成的正确率;最后针对十个标准数据集进行了实验研究,结果表明集成算法DBNNE在小规模数据集上优于Bagging及AdaBoost集成方法,而在较大数据集上也不逊色于这两种集成方法.

An Approach to Improving Diversity of Neural Network Ensemble

Ensemble learning has become one of research fields of machine learning,it dramatically improves generalization performance of classifier.After analyzing ensemble approach to both Bagging and Adaboost,we point out their some flaws.Then we present a novel approach to neural network ensemble,called DBNNE below.In this method,a diverse data set is generated to increase ensemble diversity.Moreover,to ensure high accuracy of ensemble,we test performance of ensemble when a classifier is added to ensemble .Finally,we experiment on ten representative data sets.The results show that DBNNE achieves higher predictive accuracy than Bagging and AdaBoost on small data sets and comparable performance on larger data sets.