基于低秩表示的标记分布学习算法  

Label Distribution Learning Method Based on Low-Rank Representation

作　　者：刘睿馨 刘新媛 李晨[1] LIU Ruixin;LIU Xinyuan;LI Chen(School of Software Engineering,Xi′an Jiaotong University,Xi′an 710049)

机构地区：[1]西安交通大学软件学院,西安710049

出　　处：《模式识别与人工智能》2021年第2期146-156,共11页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.61573273)资助。

摘　　要：针对标记分布学习算法忽略标记相关性信息及数据存在异常和噪声值的情况,文中提出基于低秩表示的标记分布学习算法(LDL-LRR).利用特征空间的基线性表示样本信息,实现对原始特征空间数据的降维.将低轶表示(LRR)迁移至标记空间,对模型施加低秩约束,把握数据的全局结构.分别使用增广拉格朗日乘子法和拟牛顿法求解LRR和目标函数,再通过最大熵模型预测标记分布.在10个数据集上的对比实验表明,LDL-LRR性能良好,效果稳定.Label correlations,noises and corruptions are ignored in label distribution learning algorithms.Aiming at this problem,a label distribution learning method based on low-rank representation(LDL-LRR)is proposed.The base of the feature space is leveraged to represent the sample information,and consequently dimensionality reduction of the data in the original feature space is achieved.To capture the global structure of the data,low-rank representation is transferred to the label space to impose low-rank constraint to the model.Augmented Lagrange method and quasi-Newton method are employed to solve the LRR and objective function,respectively.Finally,the label distribution is predicted by the maximum entropy model.Experiments on 10 datasets show that LDL-LRR produces good performance and stable effect.

关 键 词：标记多义性 单标记学习(SLL) 多标记学习(MLL) 标记分布学习(LDL) 低秩表示(LRR) 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]