一种非负稀疏近邻表示的多标签学习算法  

A Non-Negative Sparse Neighbor Representation for Multi-Label Learning Algorithm

作　　者：陈思宝[1] 徐丹洋[1] 罗斌[1] 

机构地区：[1]f安徽大学计算机科学与技术学院,合肥2306011

出　　处：《电子科技大学学报》2015年第6期899-904,共6页Journal of University of Electronic Science and Technology of China

基　　金：国家863项目(2014AA015104);国家自然科学基金(61202228,61472002);安徽省高校自然科学研究重点项目(KJ2012A004)

摘　　要：针对训练数据中的非线性流形结构以及基于稀疏表示的多标签分类中判别信息丢失严重的问题,该文提出一种非负稀疏近邻表示的多标签学习算法。首先找到待测试样本每个标签类上的k-近邻,然后基于LASSO稀疏最小化方法,对待测试样本进行非负稀疏线性重构,得到稀疏的非负重构系数。再根据重构误差计算待测试样本对每个类别的隶属度,最后实现多标签数据分类。实验结果表明所提出的方法比经典的多标签k近邻分类(ML-KNN)和稀疏表示的多标记学习算法(ML-SRC)方法性能更优。In order to avoid the influence of the nonlinear manifold structure in training data and preserve more discriminant information in the sparse representation based multi-label learning, a new multi-label learning algorithm based on non-negative sparse neighbor representation is proposed. First of all, the k-nearest neighbors among each class are found for the test sample. Secondly, based on non-negative the least absolute shrinkage and selectionator operator （LASSO）-type sparse minimization, the test sample is non-negative linearly reconstructed by the k-nearest neighbors. Then, the membership of each class for the test sample is calculated by using the reconstruction errors. Finally, the classification is performed by ranking these memberships. A fast iterative algorithm and its corresponding analysis of converging to global minimum are provided. Experimental results of multi-label classification on several public multi-label databases show that the proposed method achieves better performances than classical ML-SRC and ML-KNN.

关 键 词：多标签学习 稀疏近邻表示 LASSO稀疏最小化 非负重构 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]