基于最优权重的神经网络集成文本分类研究  

Study of Web document classification based on best weight neural network ensembles

作　　者：周朴雄[1] 

机构地区：[1]华南理工大学电子商务学院,广州510006

出　　处：《计算机应用研究》2008年第10期2982-2983,共2页Application Research of Computers

基　　金：广东省自然科学基金资助项目(07300647)

摘　　要：将神经网络集成思想引入Web文本分类领域,提出了利用最小估计误差策略进行最优加权网络集成的方案。具体做法是根据各网络的分类性能、各网络同其他网络的相关程度给每个网络的后验概率估计赋予不同的权值,通过加权平均提高后验概率估计的准确程度,进而提高分类率。英文数据库的实验结果表明,与经典的Bayes模型、kNN模型相比,该模型具有更高的分类精度与更快的分类速度。Inspired by the ideas of neural network ensembles, this paper constructed a multi-BP neural network modeling with best weights that was based the strategy of minimum estimate error. To do this, according to the capability of classification of each network and the degree of each network related to other networks, the different weight would assign to the probability estimates of maximum a posterior （MAP）. Further, improved the accuracy of estimate and classification. The experimental results of English database demonstrate that this model hold the better accuracy and speed than the Bayes and kNN models.

关 键 词：文本分类 神经网络集成 精度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]