一种基于证据理论的多类半监督分类算法

A Multi-class Semi-Supervised Classification Algorithm Based on Evidence Theory

为了提高多类半监督分类的性能,提出了一种基于证据理论的多类协同森林算法(DSM-Co-Forest).首先,通过"多对多"模式将有标记的多类数据随机拆分为多个二类数据集,并以此训练二类基分类器;然后,利用多个基分类器同时对未标记样本进行预测,并利用证据组合算法挑选出可信度较高的未标记样本;最后,将高可信度的未标记样本加入到原训练样本中,以迭代更新其他的基分类器,从而提高分类器的整体性能.通过在一些公共数据集上进行实验,并与其他半监督分类算法进行对比,验证了所提算法的可行性和有效性.

In order to improve the performance of multi-class semi-supervised classification,a new multi-class Co-Forest algorithm named DSM-Co-Forest is proposed on the basis of D-S evidence theory.First,through MVM mode,the multi-labeled data set is randomly split into multiple binary-class data set to train the base classifiers;then,these base classifiers are used to pick out the high reliability samples from the unlabeled data set by using the evidence combination algorithm;finally,adds these selected samples to the original training set to iteratively update the base classifiers so as to improve the overall performance of the multi-class classifier.Through comparing with other semi-supervised classification algorithms on several public data sets,the feasibility and validity of the proposed algorithm are verified.