一种权衡性能与隐私保护的推荐算法  

A Recommendation Algorithm Trading off Performance and Privacy Protection

作　　者：马黛露丝 朱海萍[1,2] 田锋[1,2] 冯沛 陈妍[1,2] 计湘婷 李玉杰 MA Dailusi;ZHU Haiping;TIAN Feng;FENG Pei;CHEN Yan;JI Xiangting;LI Yujie(Faculty of Electronic and Information Engineering,Xi’an Jiaotong University,Xi’an 710049,China;Shaanxi Province Key Laboratory of Satellite and Terrestrial Network Tech.R&D,Xi’an Jiaotong University,Xi’an 710049,China;Baidu.com Times Technology(Beijing)Co.,Ltd.,Beijing 100085,China;School of Computer Science and Engineering,Nanjing University of Technology,Nanjing 210094,China)

机构地区：[1]西安交通大学电子与信息学部,西安710049 [2]西安交通大学陕西省天地网技术重点实验室,西安710049 [3]百度时代网络技术(北京)有限公司,北京100085 [4]南京理工大学计算机科学与工程学院,南京210094

出　　处：《西安交通大学学报》2021年第7期117-123,共7页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(61877048);陕西省自然科学基础研究计划资助项目(2020JM-070);百度科研合作项目。

摘　　要：针对推荐系统中的隐私保护问题,提出一种隐私保护参数与推荐精度的均衡优化模型。以在线学习资源推荐系统为例建立矩阵分解模型,分别在数据输入和模型训练两个模块引入差分隐私噪声扰动,研究隐私保护参数ε与推荐精度的关系,并针对在线学习推荐系统数据的隐式反馈特点,提出基于资源热度负采样算法。基于西安交通大学网络学院数据集,利用百度飞桨深度学习平台,使用均方根误差作为推荐精度评价指标进行实验,结果表明:对数据基于资源热度负采样后,推荐精度更高;推荐精度与ε的倒数成正比关系,输入扰动和模型扰动分别在均方根误差取值不高于2.0和1.3且ε取值7和3时均衡效果最优;在相同ε下,当ε≤5时,模型扰动的差分隐私推荐算法推荐精度高于输入扰动的差分隐私推荐算法。Aiming at the problem of privacy protection in recommender systems,a balanced optimization model between privacy protection parameters and recommendation accuracy is proposed.Taking online learning resource recommender system as an example,this paper builds a matrix factorization model and studies the relationship between privacy protection parameters and recommendation accuracy after introducing differential privacy noise into data input module and model training module.According to the implicit feedback characteristics of online learning resource recommender system data,a negative sampling algorithm by resource-popularity is proposed.The experiment is based on the original and balanced data from the Network College of Xi’an Jiaotong University using Baidu PaddlePaddle platform,and the root mean square error is used as an evaluation index to measure the recommendation accuracy.The results show that negative sampling makes higher prediction accuracy than the original.And the recommendation prediction accuracy is directly proportional to the reciprocal of differential privacy protection parameter.When RMSEs are less than 2.0 and 1.3 and the privacy protection parameter is 7 and 3,respectively,both the input-based algorithm and the model-based algorithm achieve best balance.Moreover,when the privacy protection parameter is no more than 5,the model-based algorithm has a higher recommendation accuracy than the input-based algorithm.

关 键 词：推荐系统 差分隐私保护 矩阵分解 负采样 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]