计算机工程 ›› 2018, Vol. 44 ›› Issue (11): 300-305,312.doi: 10.19678/j.issn.1000-3428.0048759

• 开发研究与工程应用 • 上一篇    下一篇

PMUS-HOSGD张量分解方法及其在标签推荐中的应用

杨林a,b,顾军华a,b,官磊a,b,张宇娟a,b,彭玉青a,b   

  1. 河北工业大学 a.计算机科学与软件学院; b.河北省大数据重点实验室,天津 300400
  • 收稿日期:2017-09-21 出版日期:2018-11-15 发布日期:2018-11-15
  • 作者简介:杨林(1991—),男,硕士研究生,主研方向为数据挖掘、推荐系统;顾军华,教授、博士;官磊、张宇娟,硕士研究生;彭玉青,教授。
  • 基金项目:

    河北省科技计划项目(17210305D,15210345);天津市科技计划项目(16ZXHLSF00230)

PMUS-HOSGD tensor decomposition method and its application in tag recommendation

YANG Lina,b,GU Junhuaa,b,GUAN Leia,b,ZHANG Yujuana,b,PENG Yuqinga,b   

  1. a.School of Computer Science and Software; b.Big Data Key Laboratory of Hebei Province, Hebei University of Technology,Tianjin 300400,China
  • Received:2017-09-21 Online:2018-11-15 Published:2018-11-15

摘要:

目前的标签推荐系统使用张量来存储“用户-资源-标签”三维数据,以挖掘三者之间潜在的语义关联。为更好地解决三维数据的稀疏性问题,避免张量填充造成的数据失真,提出基于标签惩罚机制的张量构建方法PMUS和基于随机梯度下降的张量分解方法HOSGD。利用标签惩罚机制和用户评分构建张量,使用随机梯度下降法对张量的展开矩阵进行分解。在此基础上,结合PMUS和HOSGD提出PMUS-HOSGD算法对数据进行处理,根据结果为用户进行个性化标签推荐。在数据集MovieLens上的实验结果表明,与CubeALS、HOSVD和CF算法相比,该算法能够有效提高标签推荐的准确率。

关键词: 标签推荐, 数据稀疏性, 张量构建, 张量分解, 惩罚机制, 随机梯度下降

Abstract: The current label recommended system uses tensors to store “user-resource-label” three-dimensional data,mining the potential semantic association among the three.In order to solve the sparseness problem of three-dimensional data and avoid the data distortion caused by the tensor filling process,this paper proposes a tensor construction method based on tag punishment mechanism PMUS and a tensor decomposing method based on stochastic gradient descent HOSGD.Tensor is constructed by combining tag punishment mechanism and user score,the expansion matrix of tensor is decomposed by the method of stochastic gradient descent.On this basis,combined with PMUS and HOSGD,the PMUS-HOSGD algorithm is proposed to deal with the data,and the personalized tag recommendation is made for the user according to the results.The experimental results on the data set MovieLens show that,compared with CubeALS,HOSVD and CF algorithms,this algorithm can effectively improve the accuracy of tag recommendation.

Key words: tag recommendation, data sparseness, tensor construction, tensor decomposition, penalty mechanism, stochastic gradient descent

中图分类号: