一种基于正则化地半监督多标记学习算法.PDFVIP

第x 卷 计 算 机 研 究 与 发 展 Vol.x, Suppl. 20xx 年xx 月 JOURNAL OF COMPUTER RESEARCH AND DEVELOPMENT Aug. 20xx 一种基于正则化的半监督多标记学习算法 李宇峰 黄圣君 周志华 南京大学计算机软件新技术国家重点实验室，南京 210093 {liyf, huangsj, zhouzh}@ 摘要 本文研究半监督多标记学习问题，其中每个训练样本可隶属于多个类标记并且大部分训练样本是未标记的。这类学习问 题广泛存在于多标记学习的实际应用中，然而现有的大多数多标记学习算法只适用于传统监督学习框架。为解决该学习问题， 本文提出了一种基于正则化的半监督多标记学习方法——MASS 方法，试图利用丰富的未标记样本与标记的关系信息以帮助提 高性能。具体而言，在最小化经验风险的基础上，MASS 方法引入两种正则项分别用于描述多个标记的共性信息和约束相似样 本应拥有相似的结构化多标记输出。本文提供一个快速交替优化算法取得整体凸优化问题的全局最优解。在网页分类和基因功 能分析问题上的实验结果验证了本文方法的有效性。 关键词 多标记学习；半监督学习 中图法分类号 TP391 Regularized Semi-Supervised Multi-Label Learning Li Yu-Feng, Huang Sheng-Jun, Zhou Zhi-Hua National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210093 Abstract In this paper, we study the semi-supervised multi-label problem where each training example could be associated with multi- ple class labels while most of training examples are unlabeled. This scenario occurs in many real-world multi-label applications while most previous studies on multi-label learning typically work on traditional supervised learning setting. To address this problem, we pro- pose a regularized Multi-lAbel Semi-Supervised learning (MASS) method which exploits both the abundant unlabeled examples and the label relationships simultaneously to help improve the performance. Specifically, besides the empirical risk minimization, MASS em- ploys two regularizers to characterize the commonness among multiple labels and enforce similar instances share with similar structural multi-label outputs. Moreover, we use an efficient alternative method to obtain the optimal solution for the overall convex optimization problem. Experimental results on web page categorization and gene functional analysis demonstrate the effectiveness of our method. Keywords multi-label learning; semi-supervised learning