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。
Multi-label Specific Features Learning Algorithm Based on Label Correlation
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 gen-erated label-specific features could not comprehensively and accurately characterize the label information.To weaken this shortcom-ing,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;fur-thermore,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 multila-bel 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,respec-tively,demonstrating the effectiveness of the proposed algorithm.
国家科技期刊平台
NSTL
万方数据
