基于稀疏正则化的加权叠加集成多标签分类  

WEIGHTED SUPERPOSITION ENSEMBLE MULTIPLE LABEL CLASSIFICATION BASED ON SPARSE REGULARIZATION

作　　者：肖建芳[1] 刘缅芳[2] Xiao Jianfang;Liu Mianfang(Shantou Polytechnic,Shantou 515041,Guangdong,China;School of Mathematics and Computational Science,Hunan University of Science and Technology,Xiangtan 411100,Hunan,China)

机构地区：[1]汕头职业技术学院,广东汕头515041 [2]湖南科技大学数学与计算机科学学院,湖南湘潭411100

出　　处：《计算机应用与软件》2024年第5期286-297,共12页Computer Applications and Software

基　　金：广东省高等学校科研项目(2018GkQNCX150);2020年度广东省普通高校特色创新和青年创新人才项目立项(2020KQNCX209)。

摘　　要：为了充分挖掘成对标签的相关性以及分类器权重与分类器选择之间的关系,提出一种基于稀疏正则化的加权叠加集成多标签分类方法。提出一个稀疏正则化的加权叠加集成模型,以便于多标签分类器的选择和集成成员的构建。利用分类器权值和标签相关性来提高分类性能。进一步提出基于加速近端梯度和块坐标下降技术的优化算法来有效地获得最优解。在多个数据集上的实验结果表明,该方法能够有效实现较高精度的多标签分类。In order to fully exploit the correlation of paired labels and the relationship between classifier weight and classifier selection,a weighted superposition ensemble multiple label classification method based on sparse regularization is proposed.A sparse regularized weighted superposition ensemble model was proposed to facilitate the selection of multiple label classifiers and the construction of ensemble members.The classifier weight and label correlation were used to improve the classification performance.An optimization algorithm based on accelerated proximal gradient and block coordinate descent technique was proposed to obtain the optimal solution effectively.Experimental results on several data sets show that the proposed method can effectively achieve high precision multiple label classification.

关 键 词：多标签分类 相关性 稀疏正则化 权值 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]