Abstract:

The performance of semi-supervised learning algorithm can be enhanced by incorporating the structural information of the dataset. Based on this assumption, a novel graph-based semi-supervised learning method is proposed. A local scale scheme is presented to define different scale parameters for edge weights in different density regions. The dataset is classified by a semi-supervised learning algorithm named graph Laplacian kernels. Experiments on several datasets show that the proposed method outperforms other kernel-based semi-supervised learning algorithms.

Key words: semi-supervised learning, local scale transformation, Laplacian kernels, classification learning

摘要：

采用数据点的结构信息可以提高半监督学习的性能。为此，提出一种基于图的半监督学习方法。利用局部尺度转换对不同密度区域中的边权重设置不同的尺度参数，在此基础上构造图的拉普拉斯核分类器进行分类学习。在多个数据集上的实验显示该方法优于其他基于核的半监督分类方法。

关键词: 半监督学习, 局部尺度转换, 拉普拉斯核, 分类学习

CLC Number: 

• TP18