电子学报 ›› 2020, Vol. 48 ›› Issue (12): 2345-2351.DOI: 10.3969/j.issn.0372-2112.2020.12.009

• 学术论文 • 上一篇    下一篇

一种基于全局和局部标记相关性的多标记分类算法

朱赛赛, 贾修一, 李泽超   

  1. 南京理工大学计算机科学与工程学院, 江苏南京 210094
  • 收稿日期:2019-09-23 修回日期:2020-05-02 出版日期:2020-12-25
    • 通讯作者:
    • 贾修一
    • 作者简介:
    • 朱赛赛 男,1994年出生于安徽濉溪,硕士研究生,主要研究方向为机器学习,数据挖掘.E-mail:zhusaisworld@163.com;李泽超 男,1985年出生于河南开封,教授,博士生导师,主要研究方向为图像视频智能分析.E-mail:zechao.li@njust.edu.cn
    • 基金资助:
    • 国家重点研发计划资助 (No.2019YFB1707300); 国家自然科学基金项目 (No.61773208,No.61772275); 江苏省自然科学基金项目 (No.BK20191287); 中央高校基本科研业务费专项资金 (No.30920021131)

Exploiting Global and Local Label Correlations for Multi-label Classification

ZHU Sai-sai, JIA Xiu-yi, LI Ze-chao   

  1. School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, Jiangsu 210094, China
  • Received:2019-09-23 Revised:2020-05-02 Online:2020-12-25 Published:2020-12-25
    • Corresponding author:
    • JIA Xiu-yi

摘要: 多标记学习用于处理一个示例同时与多个类别标记相关的问题.在多标记学习中,标记相关性能够显著提升学习算法的性能.大多数现有的多标记学习算法在利用标记的相关性时,要么只使用被所有示例所共享的全局标记相关性,要么就使用局部标记相关性,它们认为不同簇中的示例应该存在不同的标记相关性.本文中,我们提出了一种同时利用全局和局部标记相关性的多标记学习算法,从而为学习进程提供更全面的标记信息.在计算全局和局部标记相关性时,我们使用了余弦相似性来获取不同标记之间的正相关性和负相关性,这样有助于我们进一步实现更可靠的多标记学习.我们在多种类型的数据集上进行了广泛的对比实验来验证所提算法的有效性.实验结果表明,该算法显著优于大多数对比算法,展现出其在多标记学习中的突出性能.

 

关键词: 多标记学习, 标记相关性, 余弦相似性

Abstract: Multi-label learning deals with the problem where each instance has a set of class labels simultaneously. In multi-label learning, label correlations have shown promising strength in improving multi-label learning. Most of the existing multi-label learning algorithms exploited either global label correlations shared among all instances, or local label correlations varied across different clusters of instances. In this study, we propose a novel multi-label learning method by simultaneously taking into account both the global and local label correlations to capture more comprehensive label information during the learning process. To calculating global and local label correlations, we utilize cosine similarity to obtain positive and negative correlations between different labels, which helps us to further achieve more reliable multi-label learning. We implemented extensive experimental comparisons based on various data sets to validate the effectiveness of our algorithm. The experimental results show that the proposed algorithm significantly outperforms most of the state-of-the art approaches, demonstrating its prominent performance for multi-label learning.

 

Key words: multi-label learning, label correlations, cosine similarity

中图分类号: