δ-距离及其在半监督增强应用.pdfVIP

第十五届全国图像图形学学术会议 -距离及其在半监督增强中的应用 王倩影， 冯国灿 ， 汤鑫 中山大学数学与计算科学学院，广州 510275 摘 要：在半监督增强学习中，许多算法都要用到测度来度量样本间的相似关系。但目前大部分的增强学习都局 限于使用欧氏距离来度量样本间的相似度。本文提出了一种新的度量： -距离。 -距离是基于测地距离但包含了 样本间隔的新度量。最直观的也是最常见的数据度量是欧氏距离，但是欧氏距离只会考虑空间中两点分开的程度， 而不考虑其他信息，比如，子空间低维流形的信息，低维流形的信息在我们现实生活中处理的高维数据起着很重 要的作用。测地距离考虑了流形的信息，但没有考虑到数据的间隔信息[1]，而此信息在半监督增强学习中又有着 很重要的作用。因此我们所提出了δ-距离，他继承了测地距离的优点[2]，同时根据数据在 邻域下的连通性建图 体现了数据的间隔信息。同时文章中还给出了计算 -距离的自适应算法。实验证明在半监督增强学习中应用 -距 离，分类器的分类性能有了很大的提高。 关键词 ： δ-距离 ; 流形; 间隔; 信息； 半监督学习 -distance and its Application in Semi-supervised Boosting WANG QIAN YING, FENG GUO CAN, TANG XIN College of Mathematics and Computational Science, Sun Yat- Sen University, Guangzhou, 510275 Abstract: In semi-supervised boosting, a metric is demanded by many algorithm to measure the similarity. But the existing methods are limited to using Euclidean distance to measure the similarity between samples. This paper presents a new metric: -distance. -distance is based on geodesic distance but incorporates the margin information of the data. The most intuitive and common metric is Euclidean distance, but it only computes the distance separating two points of the space, ignores other information, such as the sub-manifold information, which plays an important role in dealing with high-dimensional data in our real life. The geodesic distance is along the manifold which will reflect the manifold information, but it ignore the margin message of the data, which is important in semi-supervised learning. So the δ -distance is proposed. It not only inherits the benefit of geodesic distance, but also regards the margin information. An adaptive algorithm for comput