基于特征提取的选择性神经网络集成方法  

Feature Extraction-based Selective Neural Network Ensemble Method

作　　者：朱帮助[1] 

机构地区：[1]北京航空航天大学经济管理学院

出　　处：《计算机科学》2008年第3期132-133,172,共3页Computer Science

基　　金：国家自然科学基金(70471074);广东省科技计划(2004B36001051)

摘　　要：针对现有神经网络集成研究方法在输入属性、集成方式和集成形式上的不足,提出一种基于特征提取的选择性神经网络集成模型——NSNNEIPCABag。该模型通过Bagging算法产生若干训练子集;利用改进的主成分分析(IPCA)提取主成分作为输入来训练个体网络;采用IPCA从所有个体网络中选择出部分线性无关的个体网络;采用神经网络对选择出来的个体网络进行非线性集成。为检验该模型的有效性,将其用于时间序列预测,结果表明本文提出的方法的泛化能力优于流行的其它集成方法。In order to solve the problems of the existing research methods in input attribute, ensemble fashion and ensemble form, a novel selective ＆ nonlinear neural network ensemble method, i.e. NSNNEIPCABag, is proposed in this paper. In this model, some different training subsets are first generated by Bagging algorithm. Then the feature extraction technique, improved principal component analysis （IPCA）, is used to extract the data feature for training in- dividual networks, and to select the appropriate number of ensemble members from the available networks. Finally, the selected members are aggregated into an ensemble model by neural network. For illustration and testing purposes, the proposed ensemble model is applied to time series forecasting with the favor results obtained, which shows that the generalization ability of the proposed method can be superior to those of other neural network ensembles.

关 键 词：集成学习 神经网络集成 非线性集成 选择性集成 特征提取 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP393[自动化与计算机技术—计算机科学与技术]