摘要
介绍了概率神经网络(PNN)方法及其分类机理,构造了用于识别两类模式样本的PNN结构,用来对我国2000年106家上市公司进行两类模式分类.仿真结果表明,PNN对训练样本有很高的分类准确率,能达到100%;但对测试样本的分类准确率却很低,只达到69 77%.因而使总体的分类效果偏差,分类准确率只达到87 74%.进一步的仿真结果表明,该方法对我国2001年公布的13家预亏公司进行预警分析时,准确率只达到69 23%.所以PNN不太适合用来对新样本的识别和预警研究.研究结果还表明,PNN在分类效果上不如MLP(对相同的样本,多层感知器分类准确率达到98 11%),但和Yang等的PNN分类效果(分类准确率最高达到74%)相比,该文给出的PNN结构其分类效果更好.所以作为一种方法上的探讨,PNN仍不失其研究的价值.
The article introduces the method of probabilistic neural network (PNN) and its classifying principle. It constructs a PNN structure for identified two patterns samples. The PNN structure is used to separate 106 listed companies of our country in 2000 into two groups. The simulations show that, the classification accuracy rate of PNN to the training samples is very high which is up to 100%, but the classification accuracy rate of PNN to the testing samples is very low which is only 69.77%. Therefore, the classification effect to the population tends to bad and the accuracy rate is only 87.74%. Further simulating results show the predicting accuracy rate is only 69.23% when the PNN is used to predict 13 pre-distressed companies which are published in advance from China in 2001. Therefore, PNN is not suitable to identify a new sample or to carry out predicting study. The research also shows that, PNN is not as good as MLP (to the same data, the classification accuracy rate of the multilayer perceptron is 98.11%). But compare with Yang's work about PNN's classification (the classification accuracy rate is 74%) effect, the classification effect of the PNN structure given by here is better. Therefore, as a discussion of method, PNN still have research value.
出处
《系统工程理论与实践》
EI
CSCD
北大核心
2005年第5期43-48,共6页
Systems Engineering-Theory & Practice
基金
广东省自然科学基金(31906)
广东省科技厅(2004B10101033)
广州市科技局攻关项目(2004Z3 D0231)
关键词
概率神经网络
信用评价模型
模式分类
财务预警
probabilistic neural network
credit scoring model
patterns classification
financial predicting
