一种基于MDL度量的选择性扩展贝叶斯分类器

朴素贝叶斯分类器是一种简单而高效的分类器,但它的条件独立性假设使其无法表示属性间的依赖关系。TAN分类器按照一定的结构限制,通过添加扩展弧的方式扩展朴素贝叶斯分类器的结构。在TAN分类器中,类变量是每一个属性变量的父结点,但有些属性的存在降低了它分类的正确率。文中提出一种基于MDL度量的选择性扩展贝叶斯分类器(SANC),通过MDL度量,删除影响分类性能的属性变量和扩展弧。实验结果表明,与NBC和TANC相比,SANC具有较高的分类正确率。

A Selective Augmented Naive Bayesian Classifier Based on MDL Score

Naive Bayesian classifier is a simple and effective classifier,but its conditional independence assumption makes it unable to express the dependence among features.TAN classifier extends the structure of Naive Bayes classifier by adding augmenting arcs that obey certain structural restrictions.In TAN classifier,all features are constrained to have the class variable as a parent,but some features degrade its classification accuracy.The present paper presents SANC(A Selective Augmented Naive Bayesian Classifier based on MDL score) that removes features and augmenting arcs which affect the performance of classification by MDL score.Compared with NBC and TANC,experimental results show SANC has higher accuracy.