基于用户行为序列的概率矩阵分解推荐算法

Probability Matrix Factorization Recommendation Algorithm Based on User Behavior Sequence

传统的概率矩阵分解在推荐算法中取得了一定的效果,但是仍然面临数据稀疏性问题,并且对数据的利用效率不高,不能根据已有数据准确计算用户(物品)之间的关系,评分预测准确性仍然有待提高.本文利用用户对物品的评分序列信息充分挖掘用户(物品)之间的相似度关系,提出了基于用户行为序列的概率矩阵分解推荐算法UBS-PMF(Probability Matrix Factorization Recommendation Algorithm Based on User Behavior Sequence).首先根据用户对物品的评分序列和物品标签信息计算用户对标签的评分序列,即为用户的偏好转移序列,根据该序列可以计算出用户之间的相似度矩阵,同时,用户对物品的评分序列也隐藏着物品之间的关系,利用多个用户对物品的评分序列可以得到物品相似度矩阵,将所得用户(物品)相似度矩阵融入概率矩阵分解模型中进行评分预测,Movielens数据集中的实验表明该算法具有显著的效果,在评分预测准确性方面优于传统的推荐算法.

The traditional probability matrix factorization has achieved certain effects in the recommendation algorithm,but it still faces the problem of data sparsity,and the efficiency of data utilization is not high.It is impossible to accurately calculate the relationship between users(items)based on existing data,and score prediction.Accuracy still needs to be improved.In this paper,the user’s scoring sequence information of the item is used to fully exploit the similarity relationship between users(items),and a probability matrix factorization recommendation algorithm based on user behavior sequence is proposed.Firstly,according to the user’s scoring sequence of the item and the item tag information,the user’s scoring sequence of the tag is calculated,that is,the user’s preference transfer sequence,according to which the similarity matrix between the users can be calculated,and at the same time,the user’s scoring sequence for the item The relationship between the items is also hidden.The score similarity matrix of the plurality of users can be used to obtain the similarity matrix of the items,and the obtained user(item)similarity matrix is??integrated into the probability matrix factorization model for scoring prediction,and the experiment in the Movielens data set shows The algorithm has significant effects and is superior to the traditional recommendation algorithm in scoring prediction accuracy.