基于标记相关性的多标记专属特征学习算法  

Multi-label Specific Features Learning Algorithm Based on Label Correlation

作　　者：李华 王志杰 LI Hua;WANG Zhijie(Department of Mathematics and Physics,Shijiazhuang Tiedao University,Shijiazhuang 050043,China)

机构地区：[1]石家庄铁道大学数理系,河北石家庄050043

出　　处：《山西大学学报（自然科学版）》2024年第1期59-68,共10页Journal of Shanxi University(Natural Science Edition)

基　　金：国家自然科学基金(61806133)。

摘　　要：基于双标记专属特征的多标记分类算法(BILAS)是一种代表性的多标记学习算法,然而其只考虑了在标记对下取值不同的样本,忽略了取值相同的样本,使得生成的专属特征不能全面准确地刻画标记信息。针对这一不足,基于标记的二阶相关性,对标记对的全部类型样本生成专属特征,提出基于标记相关性的多标记专属特征学习算法。首先,利用基于距离的原型学习方法选择所有标记对的原型,并进一步生成相应的专属特征;然后利用标记幂集的思想构造多标记分类器。在来自MULAN(a Java library for multi-label learning)的5个公开测试数据集上进行实验,与BILAS算法和基于校准标记排序的多标记分类算法(CLR)相比,所提算法在5种多标记评价指标上综合平均排名均是第一,且分别比BILAS和CLR算法提高了20.4%和37.1%,表明了所提算法具有较好的性能。Bilabel-specific features for multi-label classification algorithm(BILAS)is a representative multi-label learning algorithm.However,it only considers samples with different values for the label pair,and ignores samples with the same value,so that the generated label-specific features could not comprehensively and accurately characterize the label information.To weaken this shortcoming,based on the second-order correlation of labels,label-specific features are generated for all types of samples of the label pair,and a multi-label specific features learning algorithm based on label correlation is proposed.Firstly,the distance-based prototype learning method is used to select prototypes of all label pairs,and then the corresponding label-specific features are generated;furthermore,using the idea of label powerset,a multi-label classifier is constructed.Experimental results on five publicly available test datasets from MULAN(a Java library for multi-label learning)show that the proposed algorithm,compared to BILAS and multilabel classification algorithm via calibrated label ranking(CLR),ranks first in terms of the comprehensive average ranking on the five multi-label evaluation metrics.Furthermore,it achieves improvements of 20.4%and 37.1%compared to BILAS and CLR,respectively,demonstrating the effectiveness of the proposed algorithm.

关 键 词：多标记学习 数据降维 相似度 原型学习 标记幂集 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]