代价敏感的直推式支持向量机算法

Transductive Support Vector Machine Based on Cost Sensitivity

针对不均衡数据集下分类超平面偏移导致直推式支持向量机样本标记准确率低的问题,结合动态代价和TSVM算法,提出一种代价敏感的TSVM算法,该算法依据类样本的空间分布信息计算类错分代价,利用KKT条件选择对当前分类超平面有影响的样本加入下一轮迭代,该算法可以提高初始分类器的分类性能,减少错误的传递和累积,从而提高标注准确率,增强最后分类器的泛化性能.最后在UCI数据集上的实验结果表明该算法在不均衡数据集下的有效性.

Under imbalanced datasets,classification hyperplane offset causes a higher classification error rate andaffects the accuracy of sample marking. A transductive support vector machine algorithm based on cost sensitivity isimproved. This method calculates the misclassification cost according to the sample spatial distribution information,and selects the samples to join the next iteration according to the KKT conditions. It can reduce the accumulationand transmission of errors and improve the accuracy of labeling. Finally,experiment results with UCI dataset showthat this method has higher detection accuracy for imbalanced datasets.