基于AUC方法评估多类别贝叶斯分类器的性能

分类器评估一般采用准确性评估.理论证明,基于AUC方法评估分类器优于准确性评估方法,但该方法局限于二类分类问题.提出一种将二类分类问题推广到多类分类问题的新方法,用纠错输出码转换得到转换矩阵,通过转换矩阵把多类分类问题转换成二类分类问题,计算二类分类的平均值来评估分类器的性能.新方法在MBNC实验平台下编程实现,并评估贝叶斯分类器的性能,实验结果表明,这种方法是有效的.

Evaluating performance of multiple Bayes classifier based on AUC method

The evaluation of classifiers has been an important study in data mining and machine learning field. AUC (area under the receiver operating characteristic curve) is determined as a better way to evaluate classifiers than predictive accuracy. However, AUC only is used for two classes to date. A new method is referred. A conversion matrix is received by using error correcting output codes. Based on conversation matrix, multiple-classifier is turned into two-classifier. Computing the AUC value for each two-classifier and average all of the AUCs of two-classifier. The average of AUC value is used as a criterion for evaluating the performance of classifiers. Making experiment in MBNC experiment platform, the results show that the new method is effective.