电子学报 ›› 2018, Vol. 46 ›› Issue (3): 569-574.DOI: 10.3969/j.issn.0372-2112.2018.03.008

• 学术论文 • 上一篇    下一篇

一种基于动态角色标识和张量分解的推荐模型

肖云鹏, 刘晏驰, 刘红, 刘媛妮   

  1. 重庆邮电大学网络与信息安全技术重庆市工程实验室, 重庆 400065
  • 收稿日期:2017-03-20 修回日期:2017-10-10 出版日期:2018-03-25
    • 作者简介:
    • 肖云鹏,男,1979年生于安徽蚌埠,现为重庆邮电大学副教授、硕士生导师.主要研究方向为机器学习、社交网络.E-mail:xiaoyp@cqupt.edu.cn;刘晏驰,男,1993年生于重庆北碚,现为重庆邮电大学硕士研究生.主要研究方向为推荐系统、机器学习.E-mail:liuyanchi030@sina.cn;刘红,女,1981年生于重庆巴南,现为重庆邮电大学副教授、硕士生导师.主要研究方向为图像安全、图像搜索.E-mail:liuhong1@cqupt.edu.cn;刘媛妮,女,1982年生于河南邓州,现为重庆邮电大学副教授.主要研究方向为移动群智感知网络、物联网安全.E-mail:liuyn@cqupt.edu.cn
    • 基金资助:
    • 国家重点基础发展研究计划 (No.2013CB329606); 国家自然科学基金 (No.61772098); 重庆市基础科学与前沿研究项目 (No.cstc2017jcyjAX0099)

A Recommendation Model Based on Dynamic Role Identification and Tensor Decomposition

XIAO Yun-peng, LIU Yan-chi, LIU Hong, LIU Yuan-ni   

  1. Chongqing Engineering Laboratory of Internet and Information Security, Chongqing University of Posts and Telecommunications, Chongqing 400065, China
  • Received:2017-03-20 Revised:2017-10-10 Online:2018-03-25 Published:2018-03-25
    • Supported by:
    • National Program on Key Basic Development Research Project of China (No.2013CB329606); National Natural Science Foundation of China (No.61772098); Basic Science and Frontier Research Project of Chongqing Municipality (No.cstc2017jcyjAX0099)

摘要: 社交网络推荐中,通常未依据用户兴趣变化进行用户角色动态标注,会造成推荐预测误差,并且用户评分数据稀疏造成评分预测不准确.根据以上问题,本文提出一种基于动态角色标识和张量分解的推荐模型.首先,针对用户角色无差别标识问题,引入信息熵指标度量用户兴趣多样性,对目标用户进行角色定量标识.其次,考虑到用户兴趣漂移现象,提出基于时间窗的动态角色标识方式,解决静态角色标识产生的个体评分数据无偏好差异问题,实现用户评分数据层次化处理.最后,为提高评分预测准确率,通过引入张量分解在数据维度转换和数据压缩的特性,构建基于"用户-项目-角色"张量分解的评分预测模型.同时,在张量分解的过程中,通过对数据缺失值处理,提高评分预测精度,生成目标用户推荐列表.实验表明,该方法缓解了用户无角色差异形成的预测误差问题,并能够有效改善评分数据稀疏情况下传统方法推荐精度不足的问题,提高推荐效率.

关键词: 推荐系统, 动态角色标识, 张量分解, 社交网络, 兴趣漂移, 稀疏性

Abstract: In the current social recommendation system, user roles are typically not dynamically annotated based on changed user interest. The flaw may lead to the prediction inaccuracy of recommendation. Besides, sparsity of user rating data can also cause imprecise prediction. According to the above problems, this paper proposes a recommendation model based on dynamic role identification and tensor decomposition. Firstly, when user roles are quantitatively identified, information entropy is used to capture the diversity of user interest for solving the problem of indiscriminate user role identification. Secondly, considering user interest drifting, the dynamic role identification method based on time window is proposed, which enables the preference difference of individual rating data generated by static role identification and realizes the hierarchical processing of user rating data. Finally, a rating prediction model based on "user-item-role" tensor decomposition is constructed. And the characteristics of tensor in data dimension transformation and data compression is introduced into the model. In addition, by dealing with the missing value, accuracy of rating prediction is improved. Experiments demonstrate that this model can alleviate inaccurate prediction caused by indiscriminate identity of user role, and can effectively improve recommendation performance compared with the traditional recommendation model.

Key words: recommendation system, dynamic role identification, tensor decomposition, social network, interest drifting, data sparsity

中图分类号: