一个基于信息论的示例学习方法

AN INFORMATION—BASED METHOD IBLE FOR LEARNING FROM EXAMPLES

本文利用信息论中信道容量、最大似然译码准则等概念,提出一个新的示例学习方法IBLE,此方法不依赖类别先验概率,特征间为强相关,具有直观的知识表示,将它用于质谱解析,结果很好,八类化合物平均正确预测率为93.96％,高于专家水平。

This paper presents a new method IBLE for learning from examples with the concepts of capacity, maximal plausible decode criterion of information theory. The method doesn't depend on the prior probability of class. In the method,the attributes are strongly associated, the knowledge representation is intelligible. We use IBLE in the interpretation of mass spectra, good result is obtained and the average predictive accuracy for eight classes of compounds is 93. 96%. This result is superior to experts.