电子学报 ›› 2017, Vol. 45 ›› Issue (9): 2057-2064.DOI: 10.3969/j.issn.0372-2112.2017.09.001

• 学术论文 •    下一篇

一种基于差分隐私和时序的推荐系统模型研究

范利云1,2, 左万利1,2, 王英1,2, 王鑫3   

  1. 1. 吉林大学计算机科学与技术学院, 吉林长春 130012;
    2. 吉林大学符号计算与知识工程教育部重点实验室, 吉林长春 130012;
    3. 长春工程学院计算机技术与工程学院, 吉林长春 130012
  • 收稿日期:2016-04-11 修回日期:2017-02-14 出版日期:2017-09-25
    • 通讯作者:
    • 左万利
    • 作者简介:
    • 范利云,女,1990年出生,河南安阳人,现为吉林大学计算机科学与技术学院硕士研究生,从事机器学习、数据挖掘等有关研究.E-mail:fanlyjlu@163.com
    • 基金资助:
    • 国家自然科学基金 (No.60973040,No.61300148,No.61602057); 吉林省科技发展计划 (No.20130206051GX,No.20130522112JH,No.20170520059JH)

Research on Recommender System Model Based on Differential Privacy and Time Series

FAN Li-yun1,2, ZUO Wan-li1,2, WANG Ying1,2, WANG Xin3   

  1. 1. College of Computer Science and Technology, Jilin University, Changchun, Jilin 130012, China;
    2. Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Changchun, Jilin 130012, China;
    3. School of Computer Technology and Engineering, Changchun Institute of Technology, Changchun, Jilin 130012, China
  • Received:2016-04-11 Revised:2017-02-14 Online:2017-09-25 Published:2017-09-25

摘要: 推荐系统的建立依赖用户的个人隐私信息,攻击者可以通过推荐的结果对用户的状态和行为进行预测.目前,虽然有对基于协同过滤近邻隐私保护的研究,但是对基于模型的隐私保护的关注度并不够高.差分隐私理论定义了一个相当严格的防攻击模型,通过添加噪声使数据失真达到隐私保护的目的,而且用户的兴趣存在兴趣漂移问题,对推荐效果造成影响,因此,提出基于差分隐私理论和时序理论构建基于模型的推荐系统.首先,根据差分隐私理论,给用户的评分数据增加小波动的符合Laplace分布的噪声,增大待分解矩阵的安全系数;然后,在随机梯度下降模型的基础上,将时序因子建模为时间权重,提高模型的准确性.实验证明该算法的准确性,并且为增强隐私研究提供了新的思路.

关键词: 推荐系统, 非负矩阵分解, 随机梯度下降法, 差分隐私, 时序理论

Abstract: Recommender system is established on users' private information.However,based on results of recommender system,attackers can predict users' states and behaviors.At present,although some researchers focus on collaborative filtering neighbor theory to preserve users' privacy,very few researchers pay enough attention to the model-based privacy-preserving.Differential privacy offers a strong degree of privacy protection by adding noise.And there is interest drift in users' interest.So this paper proposes a recommender system model based on differential privacy theory and time series theory.Firstly,according to differential privacy theory,we add some Laplace-distribution-fitted noises into users' score data to enlarge safety factor in factorization matrix.Then based on Stochastic gradient descent model,we model time series factor as time weight to improve the accuracy of the model.Experimental results demonstrate the accuracy of the algorithm,which provides a valuable perspective for privacy-preserving recommender research.

Key words: recommender system, non-negative matrix factorization, Stochastic gradient descent, differential privacy, time series

中图分类号: