基于密文的去中心化兴趣点推荐研究

DECENTRALIZED MATRIX FACTORIZATION FOR POI RECOMMENDATION BASED ON CIPHERTEXT

针对集中式矩阵分解模型在进行兴趣点推荐时存在的内存需求多、计算成本高、隐私泄露风险大等问题,提出基于密文的去中心化推荐模型CDMF。比起统一收集处理所有数据的传统矩阵分解推荐方法,CDMF保持用户的数据在个人终端,以防出现大规模的集中计算;结合随机游走和Paillier加密算法实现基于密文的去中心化推荐;为了提高算法的时间性能,CDMF进一步引入个人兴趣点集合。在Foursquare和Gowalla真实数据集上进行实验,实验结果从精确率、召回率及时间性能三方面证明了CDMF的有效性。该方法在保护隐私的同时,相比经典矩阵分解模型其精确率和召回率分别提高1百分点和9百分点。

Matrix Factorization(MF) has many defects due to its high memory requirement, expensive computation cost, huge privacy risk and so on. This paper presents a ciphertext-based decentralized MF model(CDMF) to solve these problems. Compared with the traditional MF that collects and processes all data uniformly, the model kept the user’s data in his own terminal in order to prevent large-scale centralized computing. It combined random walk and Paillier encryption algorithm for ciphertext-based decentralized MF, followed by introducing personal POI set to improve the time performance of ciphertext algorithm. Experiments on Foursquare and Gowalla real data sets demonstrate the validity of CDMF in terms of precision, recall rate and time performance, which can improve the average precision by 1 percentage point and recall rate by 9 percentage points compared with traditional MF while privacy is protected.