一种用于半监督学习的核优化设计

A Kernel Optimization Design for Semi-supervised Learning

半监督学习研究主要关注当训练数据的部分信息缺失的情况下,如何获得具有良好性能和推广能力的学习机器。本文我们提出了一种基于核优化的半监督学习框架,将数据嵌入到高维特征空间,从而与线性分类器等价。在核的设计上,采用了基于谱分解的无监督核设计,提出了学习边界,通过最小化边界来获得最优核表示。通过实验,对不同的核方法进行了比较,证明了我们结论的正确性。

Semi-supervised learning aims to obtain good performance and learning ability under lacking of some information on training examples. We proposed a semi-supervised learning framework based on optimizing kernel,which project data into high dimensional feature space and equal to linear classification. In kernel design, we applied spectral feature decomposition to unsupervised kernel design, and found optimal kernel by minimizing learning bound. With experimental results, we demonstrated our theory by comparison of different kernel approaches.