一种安全的基于分歧的半监督分类算法

为提高半监督分类的性能，提出一种安全的基于分歧的半监督分类算法Safe Co-SSC。通过有标记样本训练3个有监督分类器，利用无标记样本的信息增加分类器的差异性，采取3个分类器加权投票的策略实现对无标记样本的伪标记；对伪标记样本进行二次验证，选用能使分类器误差减小的新增标记样本扩充标记样本集。保证新样本的添加既减小了分类器的分类误差，又提高了分类器的分歧性。对UCI数据集进行分类实验的结果表明，该算法具有较高的分类率和样本标记率。  

A Safe Semi-supervised Classification Algorithm Based on Disagreement

In order to improve the performance of semi-supervised classifier , a safe disagreement-based semi-supervised classifica-tion algorithm named Safe Co-SSC was proposed .The limited labeled samples were divided into three equal training sets and used to train three classifiers by a supervised learning algorithm .A large number of unlabeled samples were used to increase the differences be-tween the classifiers and the weighted voting strategy was used to achieve pseudo -labeled for unlabeled samples .Passing through sec-ondary verification, the ones making classifier error minimum were added into the labeled samples set .Finally, the experiment was car-ried out on the UCI data set , the results showed that the proposed algorithm had higher classification rate and sample labeling rate .