基于成对约束的偏标记数据消歧算法  

Partial Label Data Disambiguation Algorithm Based on Pairwise Constraints

作　　者：征察 吉立新[1] 高超[1] 李邵梅[1] 吴翼腾 ZHENG Cha;JI Li-Xin;GAO Chao;LI Shao-Mei;WU Yi-Teng(National Digital Switching System Engineering&Techno-logical R&D Center of China,Zhengzhou 450002)

机构地区：[1]国家数字交换系统工程技术研究中心,郑州450002

出　　处：《自动化学报》2020年第7期1367-1377,共11页Acta Automatica Sinica

基　　金：国家自然科学基金(61601513)资助。

摘　　要：偏标记数据消歧是利用偏标记数据进行机器学习的基础.针对偏标记数据中广泛存在的数据不平衡问题,以及现有消歧算法对样本间约束信息利用不足的问题,本文提出一种基于成对约束的偏标记数据消歧算法.首先,基于低秩表示,推导出数据不平衡条件下样本低秩表示系数和样本相似度之间的关系;其次,基于推导结果,分别构建基于样本间正约束和负约束的图模型,通过最小化图模型的能量函数求解偏标记数据的标签.在5个公开数据集上的实验结果表明本文方法相对基准算法在消歧准确率上平均提高了2.9%~14.9%.Partial label data disambiguation is the basis of machine learning using partial label data.In order to solve the data imbalance problem widely existing in partial label data,and the problem that the existing disambiguation algorithms have insufficient utilization of constraints between samples,a partial label data disambiguation algorithm based on pairwise constraints is proposed in this paper.Firstly,the relation between low-rank representation coefficients and sample similarities in unbalanced datasets is deduced by utilizing low-rank representation.Secondly,according to the deduced results,two graphs are created based on positive constraint and negative constraint respectively.Finally,the labels of partial label data samples are obtained by minimizing energy functions based on graphs.Experimental results on five open datasets indicate that the proposed algorithm outperforms benchmark algorithms by 2.9%~14.9%at disambiguation accuracy.

关 键 词：偏标记数据 消歧 数据不平衡 低秩表示 成对约束 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]