计算机工程 ›› 2019, Vol. 45 ›› Issue (11): 172-176.doi: 10.19678/j.issn.1000-3428.0053016

• 人工智能及识别技术 • 上一篇    下一篇

基于嵌入式特征提取的多标记分类算法

王晓莹1a, 谢钧1a, 陶性留1b, 邵东生1a, 王忠2   

  1. 1. 陆军工程大学 a. 指挥控制工程学院;b. 通信工程学院, 南京 210007;
    2. 31006部队, 北京 100840
  • 收稿日期:2018-10-29 修回日期:2018-12-20 发布日期:2019-01-15
  • 作者简介:王晓莹(1994-),女,硕士研究生,主研方向为多标记学习、智能信息处理;谢钧(通信作者),教授、博士;陶性留、邵东生,硕士研究生;王忠,高级工程师。
  • 基金项目:
    国家自然科学基金(61702543)。

Multi-Label Classification Algorithm Based on Embedded Feature Extraction

WANG Xiaoying1a, XIE Jun1a, TAO Xingliu1b, SHAO Dongsheng1a, WANG Zhong2   

  1. 1a. College of Command and Control Engineering;1b. College of Communication Engineering, Army Engineering University, Nanjing 210007, China;
    2. 31006 Troops, Beijing 100840, China
  • Received:2018-10-29 Revised:2018-12-20 Published:2019-01-15

摘要: 基于单标记分类的降维及特征选择方法难以直接运用到多标记学习中,而将多标记学习问题独立分解为多个单标记学习问题再进行降维会丢失标记的相关性信息。为此,提出一种基于嵌入式特征提取的多标记分类算法,将非负矩阵分解引入到多标记学习过程中,在对原始多标记数据集进行特征提取的同时,减少冗余特征、不相关特征及高维特征对多标记分类的影响。在4个公开的标准数据集上进行对比实验,结果表明该算法能对数据进行有效降维,在准确度、精度、F度量值等评价指标上相比传统BR、CC、LM算法具有更好的分类性能。

关键词: 多标记学习, 非负矩阵分解, 特征提取, 降维, 乘性迭代, 特征转换

Abstract: Dimensionality reduction and feature selection methods based on single-label classification cannot be directly applied to multi-label learning.If a multi-label learning problem is composed into multiple independent single-label learning problems to enable dimensionality reduction,the correlation between labels will be lost.To address the problem,this paper proposes a multi-label classification algorithm based on embedded feature extraction.The algorithm introduces non-negative matrix factorization into multi-label learning.In feature extraction of original multi-label datasets,the redundant or irrelevant features are reduced and effects of high-dimensional features on multi-label classification are relieved.The comparison experiments are performed on four public standard data sets,and results show that the proposed algorithm can effectively reduce dimensions of data,and lead to better classification performance than traditional BR,CC,LM and other algorithms in terms of multiple evaluation indicators,including accuracy,precision and F measure.

Key words: multi-label learning, Non-Negative Matrix Factorization(NMF), feature extraction, dimensionality reduction, multiplicative iteration, feature transform

中图分类号: