关键词: 评价准则, 不平衡数据学习, 支持向量机, GAF准则, GBF准则

Abstract:

As most traditional classifiers optimized with the accuracy metric are unsuitable for imbalanced data learning（IDL）,this paper performs a meta-learning on a support vector machine（SVM） model,and investigates the IDL affected by such metrics as the accuracy,the balance accuracy,the geometric mean,the F1 score,the information gain,the AUC（Area Under ROC Curve）,as well as the two new metrics proposed in this paper,namely GAF and GBF.Moreover,simulation experiments are conducted on 16 imbalanced datasets from UCI,with a statistical analysis of the experimental results being also carried out.It is indicated that（1） there are distinct differences in the effects of these metrics on the classifier＇s performances;（2） even for the support vector machine（SVM）,an advanced learning method,its output classifier is still readily biased to majority class when the classifier is selected by maximizing the accuracy;（3） through the optimization with the help of other metrics,it is feasible to output bias-rectified SVM classifiers,which are of better overall performance,especially in terms of the prediction ability for minor classes;and（4） the output SVM classifiers optimized with GAF and GBF metrics are of stable and good performance.

Key words: evaluation metric, imbalanced data learning, support vector machine, GAF metric, GBF metric