融合标签局部相关性的标签分布学习  

Label distribution learning by fusion of local correlation of labels

作　　者：容斌元 徐媛媛[1] 吕亚兰 张恒汝[1] RONG Bin-yuan;XU Yuan-yuan;Lü Ya-lan;ZHANG Heng-ru(School of Computer Science,Southwest Petroleum University,Chengdu610500,Sichuan,China)

机构地区：[1]西南石油大学计算机科学学院,四川成都610500

出　　处：《山东大学学报（理学版）》2022年第7期53-64,共12页Journal of Shandong University(Natural Science)

基　　金：国家自然科学基金资助项目(61902328);中央引导地方科技发展专项资金(2021ZYD0003);南充市科技局应用基础研究项目(SXHZ040)。

摘　　要：提出了一种融合标签局部相关性的标签分布学习(label distribution learning,LDL)算法,该算法分为3个阶段。初始预测阶段构建多层神经网络模型,将样本的原始特征作为输入、初始预测的标签分布作为输出;局部矫正阶段首先利用k⁃means聚类算法获得不同类所描述的局部信息,然后针对不同类计算对应的协方差矩阵,利用该矩阵来矫正初始预测的标签分布,获得每个类对应的矫正标签分布;标签融合阶段对矫正后的标签分布进行加权,再与初始预测的标签分布进行融合,得到最终的预测分布。在8个公开数据集上与9种常用的LDL算法进行对比实验,结果表明本文的算法能较好地描述标签局部相关性,在多个主流评估指标上排名靠前。This paper proposes an LDL algorithm,which integrates the local correlation of labels.The algorithm is divided into three stages.In the initial prediction stage,this paper constructs a multi⁃layer neural network,which takes the original features as the input and the initial prediction label distribution as the outputs.In the local correction stage,first we use k⁃means to obtain the local information described by different clusters.Then,for each class,we calculate the corresponding covariance matrix,and finally use this matrix to correct the initial predicted label distribution to obtain the corrected label distribution.In the label fusion stage,the corrected label distribution is weighted,and then fused with the initial predicted label distribution to obtain the final predicted distribution.Compared with night popular LDL algorithms,experiments were conducted on eight different public datasets.The results show that our algorithm can better describe the local correlation of labels,and ranks higher on mainstream evaluation measures.

关 键 词：标签分布学习 标签相关性 标签融合 k⁃means 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]