一种基于差分隐私和时序的推荐系统模型研究

发布时间：2018-07-22 21:01

【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.
[Abstract]:The establishment of recommendation system depends on the user's personal privacy information, and the attacker can predict the user's state and behavior through the recommended results. At present, although there are some researches on the privacy protection based on cooperative filtering, the attention to model-based privacy protection is not high enough. Differential privacy theory defines a strict anti-attack model, which distorts the data to achieve the purpose of privacy protection by adding noise, and the problem of interest drift in user's interest affects the effect of recommendation. A model-based recommendation system based on differential privacy theory and time sequence theory is proposed. First of all, according to the difference privacy theory, the user's rating data is given a small fluctuation of the noise consistent with the Laplace distribution, and the safety factor of the matrix to be decomposed is increased. Then, on the basis of the stochastic gradient descent model, Time series factor is modeled as time weight to improve the accuracy of the model. Experiments show that the algorithm is accurate and provides a new way to enhance privacy.
【作者单位】：吉林大学计算机科学与技术学院;吉林大学符号计算与知识工程教育部重点实验室;长春工程学院计算机技术与工程学院;
【基金】：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)
【分类号】：TP391.3

【相似文献】