ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2018, Vol. 55 ›› Issue (8): 1751-1759.doi: 10.7544/issn1000-1239.2018.20180362

所属专题: 2018数据挖掘前沿进展专题

• 人工智能 • 上一篇    下一篇

基于神经网络探究标签依赖关系的多标签分类

宋攀,景丽萍   

  1. (交通数据分析与挖掘北京市重点实验室(北京交通大学) 北京 100044) (16120413@bjtu.edu.cn)
  • 出版日期: 2018-08-01
  • 基金资助: 
    国家自然科学基金项目(61370129,61375062,61632004,61773050) This work was supported by the National Natural Science Foundation of China (61370129, 61375062, 61632004, 61773050).

Exploiting Label Relationships in Multi-Label Classification with Neural Networks

Song Pan, Jing Liping   

  1. (Beijing Key Laboratory of Traffic Data Analysis and Mining (Beijing Jiaotong University), Beijing 100044)
  • Online: 2018-08-01

摘要: 多标签学习广泛应用于文本分类、图像标注、视频语义注释、基因功能分析等问题.最近,多标签学习受到大量的关注,成为机器学习领域中的研究热点.然而,已有的算法并不能充分地探究标签之间的依赖关系和解决标签缺失问题,为此提出一种基于神经网络探究标签依赖关系的算法NN_AD_Omega,它能够有效地处理这2个挑战.NN_AD_Omega算法在神经网络顶层加入Ω矩阵刻画标签之间的依赖关系,标签之间的依赖关系可通过充分挖掘数据内在特点得到.当实例部分标签缺失时,学到的标签之间依赖关系能够有效提高预测效果.为了高效地求解模型,采用最小批梯度下降方法(Mini-batch-GD),其中学习率的自适应计算采用AdaGrad技术.在4个标准多标签数据集上的实验结果表明,提出的算法能够探究标签之间的依赖关系和处理标签缺失问题,且其性能优于当前基于神经网络的多标签学习算法.

关键词: 神经网络, 标签依赖关系, 多标签学习, 分类, 标签缺失

Abstract: Multi-label learning is critical in many real world application domains including text classification, image annotation, video semantic annotation, gene function analysis, etc. Recently, multi-label learning has attracted intensive attention and generated a hot research topic in machine learning community. However, the existing methods do not adequately address two key challenges: exploiting correlations between labels and making up for the lack of labeled data or even missing labels. A NN_AD_Omega model via neural network for exploring labels dependencies is proposed to handle these two challenges efficiently. NN_AD_Omega model introduces an Omega matrix in the top layer of the neural network to characterize the labels dependencies. As a good by-product, the learnt label correlations have ability to improve prediction performance when the instances’ partial labels are missing because they can capture the intrinsic structure among data. In order to solve the model efficiently, we use the mini-batch gradient descent (Mini-batch-GD) method to solve the optimization problem, meanwhile, the AdaGrad technique is adopted to adaptively search the learning rate. Experiments on four real multi-label datasets demonstrate that the proposed method can exploit the label correlations and handle the missing label data, and obtain promising and better label prediction results than the state-of-the-art neural network based multi-label learning methods.

Key words: neural network, label relationships, multi-label learning, classification, missing labels

中图分类号: