摘要
一个好的核函数能提升机器学习模型的有效性,但核函数的选择并不容易,其与问题背景密切相关,且依赖于领域知识和经验。核学习是一种通过训练数据集寻找最优核函数的机器学习方法,能通过有监督学习的方式寻找到一组基核函数的最优加权组合。考虑到训练数据集获取标签的代价,提出一种基于标签传播的半监督核学习方法,该方法能够同时利用有标签数据和无标签数据进行核学习,通过半监督学习中被广泛使用的标签传播方法结合和谐函数获得数据集统一的标签分布。在UCI数据集上对提出的算法进行性能评估,结果表明该方法是有效的。
A good kernel function can improve the performance of machine learning models. However, it is not easy to properly determine a kernel since it is closely related to application background and relies on domain knowledge and experience. Kernel learning is a machine learning method which seeks an optimal kernel funetion with a given training data set. It often seeks an optimal weighted combination of a pre-defined set of base kernel functions. Considering the cost of acquiring labeled training samples,we propose a semi-supervised kernel learning method based on label propagation, which makes use of labeled and unlabeled samples simutaneously to perform kernel learning,and applies label propagation method,a popular method in semi-supervised learning, combined with harmonic ffmction to obtain a unified distribution of the whole data set. The proposed metod is evaluated on the UCI benchmark data set and the results show its effectiveness.
出处
《电脑与电信》
2013年第11期35-37,共3页
Computer & Telecommunication
基金
广东工业大学高教研究基金项目
项目编号:2013Y04
广东省大学生创新创业训练计划项目
项目编号:1184510037
广州市海珠区科技计划项目
项目编号:2011-YL-05
关键词
核学习
半监督学习
标签传播
和谐函数
支持向量机
kernel learning
semi-supervised learning
label propagation
harmonic function
support vector machine
