摘要
分析了统计判别和神经网络分类方法的特性和优缺点,提出了两种不同类型的方法集成的策略,先用神经网络自学习功能变换样本,使其分布有利于分类;再用统计方法提取特征,进而建立判别分类模型.基于该策略设计了G.T-CCA-Bayes集成方法,并应用于三个复杂模式分类问题——留兰香问题、胺类毒性问题、双螺旋问题,效果良好.对照比较表明,该集成方法适用面广,计算规范,概率意义明确,误判率低,与单一的统计分析或神经网络方法相比,有明显的优势.
Statistical discriminant and neural networks are two different pattern classification methods. After comparing their merits and demerits, a novel strategy is put forward that integrates the two methods. First, the neural network transforms the original data set. Then, the statistical method extracts from the resulting data set the characteristic components that are more preferable for classification. Finally, the discriminant analysis model is built on these components. Based on this strategy, a new integrated method, the G.T-CCA-Bayes method, is designed. Three complex pattern classification problems-spearmint essence problem, amine toxicity problem and two spirals problem-were used to estimate the classification accuracy. For comparison, another three models were built. When applied to these problems, the integrated method shows better performance than others. Comparison experiments show that the integrated method has prominent advantages over the statistical method or the neural network method. It has good adaptability and low error rate with definite meaning in probability.
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2002年第6期601-606,共6页
Journal of Zhejiang University：Engineering Science
基金
国家自然科学基金资助项目(20076041).
