计算机工程

• 人工智能及识别技术 • 上一篇    下一篇

基于Hessian正则化的多视图联合非负矩阵分解算法

王超锋,施俊,吴金杰,朱捷   

  1. (上海大学 通信与信息工程学院,上海 200444)
  • 收稿日期:2016-07-11 出版日期:2017-11-15 发布日期:2017-11-15
  • 作者简介:王超锋(1992—),男,硕士研究生,主研方向为机器学习;施俊,教授;吴金杰、朱捷,硕士研究生。
  • 基金项目:
    国家自然科学基金面上项目(61471231,81627804)。

Hessian Regularization Based Factorization Algorithm Combining Multi-view and Non-negative Matrix

WANG Chaofeng,SHI Jun,WU Jinjie,ZHU Jie   

  1. (School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China)
  • Received:2016-07-11 Online:2017-11-15 Published:2017-11-15

摘要: 非负矩阵在表征多视图数据时没有考虑数据本身的流型结构,不能有效表达数据内部信息。为此,提出一种基于Hessian正则化的非负矩阵分解算法。利用Hessian泛函的L2模,保持样本局部拓扑结构,并扩展成基于Hessian正则化的联合非负矩阵分解算法,以对多视图数据进行变换。实验结果表明,基于Hessian正则化的非负矩阵分解算法和基于Hessian正则化的联合非负矩阵分解算法的聚类精度以及互信息值都有较大提高,2种算法的数据变化性能都优于传统非负矩阵分解算法。

关键词: Hessian正则化, 回归模型, 非负矩阵分解, 多视图数据, 聚类

Abstract: Non-negative matrix does not consider the manifold of data when represents multi-view data,which results in the ineffective express of the data internal expression.In this paper,Hessian regularized Non-negative Matrix Factorization(NMF) is proposed.By using the L2 model of Hessian functional,the local topology of the sample is preserved and the algorithm is further extended into Hessian Regularized Joint Non-negative Matrix Factorization(HR-J-NMF) to work on multi-view data.Experimental results show that the Hessian regularized NMF and the HR-J-NMF have a great improvement in both clustering accuracy and mutual information value.The performance of the two algorithms is superior to that of the traditional NMF algorithm.

Key words: Hessian regularization, regression model, Non-negative Matrix Factorization(NMF), multi-view data, clustering

中图分类号: