融合时间特征的联邦矩阵分解推荐算法

Federated Matrix Factorization Recommendation Algorithm with Temporal Feature Integration

摘要: 目前矩阵分解推荐系统在集中环境下存在隐私泄露的风险，且更多的数据拥有者不愿提供自身的数据，应用于分布式环境下的联邦矩阵分解推荐系统应用而生。传统的联邦矩阵分解模型在数据稀疏的情况下推荐准确率低，没有考虑用户的兴趣随时间变化的动态性。本文针对以上问题，引入联邦矩阵分解模型与时间隐语义模型相结合，提出一种融合时间特征的联邦矩阵分解推荐算法TF-FedMF (Federated Matrix Factorization Recommendation Algorithm with Temporal Feature Integration)。该算法在联邦矩阵分解框架中加入时间特征，用于捕捉用户行为随时间变化的趋势，提高了推荐系统的时效性和准确性；同时，结合同态加密对上传的梯度信息进行加密，增强算法的安全性。通过MovieLens数据集进行实验对比，实验结果表明，所提出的算法较其它算法在兼顾用户隐私安全性的同时，具有较高的推荐准确性。

Abstract: At present, the matrix decomposition recommendation system has the risk of privacy leakage in a centralized environment, and more data owners are unwilling to provide their own data. Therefore, the application of the federated matrix decomposition recommendation system in a distributed environment has emerged. The traditional federated matrix decomposition model has low recommendation accuracy when the data is sparse, and does not consider the dynamic nature of user interests changing over time. In view of the above problems, this paper introduces the combination of the federated matrix decomposition model and the temporal latent semantic model, and proposes a federated matrix decomposition recommendation algorithm TF-FedMF (Federated Matrix Factorization Recommendation Algorithm with Temporal Feature Integration) with temporal feature integration. The algorithm adds temporal features to the framework of federated matrix decomposition to capture the trend of user behavior changing over time, thereby improving the timeliness and accuracy of the recommendation system; at the same time, the uploaded gradient information is encrypted by combining homomorphic encryption to enhance the security of the algorithm. An experimental comparison is carried out on the MovieLens dataset. The experimental results show that the proposed algorithm has higher recommendation accuracy than other algorithms while taking into account user privacy and security.