In classifier training process, the introduction of un labeled data can cause noise data, and it reduces classification accuracy. This paper proposes Confidence Estimation for Semi-supervised Learning based on graph(CESL) algorithm. The algorithm makes use of structure information of sample data to calculate classification probability of unlabeled data explicitly. Combined with multi-classifiers, the algorithm estimates the confidence of unlabeled data implicitly and improves the selection criteria. With dual-confidence estimation, the unlabeled data is selected to update classifiers. Experiments on UCi datasets prove the efficiency of this algorithm.%在分类器训练过程中,无标记数据的引入容易产生噪音,从而降低分类精度.为此,提出一种基于图的置信度估计半监督协同训练算法.利用样本数据自身的结构信息,计算无标记样本所属类别概率.采用多分类器对无标记数据进行置信度估计,以提高无标记数据挑选标准,减少噪音数据的引入.在UCI数据集上的对比实验验证了该算法的有效性.
展开▼