融合社交信息的矩阵分解改进推荐算法

矩阵分解的推荐模型具有推荐精度高和易扩展等特点，已成为目前融合社交信息构建推荐系统的主要模型，但在分解过程中，用户偏好矩阵和物品特征矩阵初始赋值的随机性影响了推荐的性能，忽略了物品以及用户之间隐含的联系与区别。为此，提出一种基于社交信息的矩阵分解改进算法。将评分值分别与社交信息和物品的特征属性相结合，构建用户相似网络与物品相似网络，同时应用社区划分充分挖掘用户、物品之间的潜在关系，并按不同类型节点的近邻差异性，通过建立核心、非核心节点的偏好向量与特征向量得到矩阵分解初始矩阵。在公开数据集上的实验结果表明，该算法的推荐性能优于MF、SR2等同类型算法，运行迭代次数明显降低。

Improved Matrix Factorization Algorithm Using Social Information for Recommendation

The recommendation models using matrix factorization have high accuracy and scalability, and they are preferred by most researchers in building a recommendation system that integrates social information.However, in the process of factorization, the random initial assignment of the user preference matrix and the item feature matrix affects the recommendation performance, as it ignores the potential connections and differences between items and users. Therefore, an improved matrix factorization algorithm based on social information is proposed.By integrating the ratings with social information and item feature attributes, the user similarity network and the item similarity network are constructed respectively.At the same time, the community division is used to fully explore the potential relationships between users and items.Then according to the difference of the neighbors between different types of nodes, the preference vector and eigenvector of the core nodes and non-core nodes are established to obtain the initial matrix of matrix decomposition.Experimental results on the public data set show that the proposed algorithm displays better recommendation performance than MF, SR2 and other algorithms, and significantly reduces the number of required iterations.