计算机工程与应用 ›› 2017, Vol. 53 ›› Issue (7): 88-96.DOI: 10.3778/j.issn.1002-8331.1510-0019

• 大数据与云计算 • 上一篇    下一篇

多源适应多标签分类框架

姚  哲1,陶剑文2   

  1. 1.诺丁汉大学 计算机科学与工程学院,英国 诺丁汉
    2.浙江大学宁波理工学院 计算机与数据工程学院,浙江 宁波 315100
  • 出版日期:2017-04-01 发布日期:2017-04-01

 Multi-source adaptation multi-label classification framework via joint sparse feature selection and shared subspace learning

YAO Zhe1, TAO Jianwen2   

  1. 1.School of Information Science and Engineering, University of Nottingham, Nottingham, UK
    2.School of Computer and Data Engineering, Ningbo Institute of Technology, Zhejiang University, Ningbo, Zhangjing 315100, China
  • Online:2017-04-01 Published:2017-04-01

摘要: 多源适应学习是一种旨在提升目标学习性能的有效机器学习方法。针对多标签视觉分类问题,基于现有的研究进展,研究提出一种新颖的联合特征选择和共享特征子空间学习的多源适应多标签分类框架,在现有的图Laplacian正则化半监督学习范式中充分考虑目标视觉特征的优化处理,多标签相关信息在共享特征子空间的嵌入,以及多个相关领域的判别信息桥接利用等多个方面,并将其融为一个统一的学习模型,理论证明了其局部最优解只需通过求解一个广义特征分解问题便可分别获得,并给出了算法实现及其收敛性定理。在两个实际的多标签视觉数据分类上分别进行深入实验分析,证实了所提框架的鲁棒有效性和优于现有相关方法的分类性能。

关键词: 特征选择, 共享特征子空间, 多源适应学习, 多标签学习

Abstract: Multi-source adaptation learning is an effective machine learning method to boost the target learning performance of interest. For multi-label visual classification applications, motivated by the advances of existing arts such as feature selection and subspace learning, this paper proposes a novel Multi-source Adaptation Multi-Label classification framework via Joint Feature Selection and shared feature Subspace learning(MAML-JFSS). Specifically, the proposed framework simultaneously considers in the classical graph Laplacian regularization semi-supervised learning scheme such important issues as the target visual feature preprocessing, the correlation among different labels encoded in some optimal shared feature subspace, and leveraging the discriminant information from multiple related source domains, and then integrates them into unified learning model. It shows by theory that the optimal solution of the proposed framework can be obtained by performing a generalized eigen-decomposition task. And then the implementing algorithm and the convergence analysis for MAML-JFSS are given. Extensive experiments on two real-world multi-label visual classification tasks are conducted. The experimental results show that the proposed algorithm is effective and can obtain comparable or even superior to related works.

Key words: feature selection, shared feature subspace, multi-source adaptation learning, multi-label learning