基于概率矩阵分解的用户相似度计算方法及推荐应用

用户相似度计算的合理性直接影响到协同过滤推荐的效果。提出了一种基于时序行为关系的用户消费网络图构建方法,通过定义用户间非对称相似度计算方法,确定用户间的初始相似度矩阵,然后利用概率矩阵分解的方法重构用户的相似度矩阵,挖掘潜在的用户近邻,将近邻关系应用到目标用户的项目推荐中,同时提出了基于时序行为关系和矩阵分解的协同过滤推荐框架结构。在实际数据集上对具体参数进行实验,并和其他方法进行了比较。实验结果表明,该方法可以有效提高协同过滤推荐效果。

User similarity calculation method based on probabilistic matrix factorization and its recommendation application

The rationality of user similarity computation directly affects the effect of collaborative filtering. A method is proposed to construct a user consumption network and a calculation formula of asymmetric user similarity was defined. The initial user similarity matrix is set up based on the user asymmetric similarity values and reconstructed via probabilistic matrix factorization. And then, potential nearest neighbor mined is applied to personalized recommendation of the target user. Meanwhile, a collaborative filtering recommendation framework is put forward based on sequential behavior and probabilistic matrix factorization. The actual parameters are experimented on real-world dataset and compared with other methods. Experimental results show that the proposed method can improve the effect of collaborative filtering recommendation.