最大规范化依赖性多标记半监督学习方法  

Normalized Dependence Maximization Multi-label Semi-supervised Learning Method

作　　者：张晨光[1] 张燕[1] 张夏欢 

机构地区：[1]海南大学信息科学技术学院,海口570228 [2]北京凌云光视公司图像处理部,北京100097

出　　处：《自动化学报》2015年第9期1577-1588,共12页Acta Automatica Sinica

基　　金：国家自然科学基金(11261015);海南省高等学校科学研究项目(Hjkj2012-01)资助~~

摘　　要：针对现有多标记学习方法大多属于有监督学习方法,而不能有效利用相对便宜且容易获得的大量未标记样本的问题,本文提出了一种新的多标记半监督学习方法,称为最大规范化依赖性多标记半监督学习方法 (Normalized dependence maximization multi-label semi-supervised learning method).该方法将已有标签作为约束条件,利用所有样本,包括已标记和未标记样本,对特征集和标签集的规范化依赖性进行估计,并以该估计值的最大化为目标,最终通过求解带边界的迹比值问题为未标记样本打上标签.与其他经典多标记学习方法在多个真实多标记数据集上的对比实验表明,本文方法可以有效从已标记和未标记样本中学习,尤其是已标记样本相对稀少时,学习效果得到了显著提高.In view of the problems that most of present multi-label learning methods are supervised learning methods and cannot effectively make use of relatively inexpensive and easily obtained large number of unlabeled samples, this paper puts forward a new multi-label semi-supervised learning method, called normalized dependence maximization multi-label semi-supervised learning method（DMMS）. The DMMS regards labeled samples as constraint conditions, estimates the normalized dependency of feature and label sets on all samples including labeled and unlabeled samples, and maximizes the estimation by finally addressing a trace ratio optimization problem with constraint conditions for label unlabeled samples. Experiments comparing DMMS with the state-of-the-art multi-label learning approaches on several real-world datasets show that the DMMS can effectively learn from labeled and unlabeled samples, especially when the labeled is relatively rare, the learning performance can be improved greatly.

关 键 词：规范化依赖性 多标记学习 半监督学习 迹比值 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]