加入用户对项目属性偏好的奇异值分解推荐算法

摘要

由于目前的矩阵分解推荐算法在解决项目冷启动问题时,没有充分利用项目的属性偏好信息与用户评分行为的交互信息,因此提出了加入用户对项目属性偏好的奇异值分解推荐(UC-SVD)算法.该算法综合考虑项目属性和用户对项目的评分,不仅在矩阵分解算法中加入了项目的属性信息,同时通过对评分数据集和属性数据集的综合分析,得出用户对项目属性的偏好矩阵,将项目属性特征因子和用户对项目属性的偏好特征因子一并加入到矩阵分解中.在数据集Movielens、HetRec2011上进行实验,结果表明,与经典矩阵分解协同过滤算法相比,所提算法不仅在一定程度上解决了项目的冷启动问题,而且在同等条件下的均方根误差平均降低了3.5％,平均绝对误差平均降低了3％,尤其是在更为稀疏的HetRec2011数据集上,项目属性对用户评分行为的影响更加明显,改进算法在推荐精度上表现出更大的优越性.%The traditional recommendation algorithm based on matrix factorization is facing the problem that in cold start of items,the existing recommendation algorithm does not make full use of the interactive information about item attributes and user's rating behavior,therefore a singular value decomposition recommendation algorithm is proposed (UC-SVD) considering the item attributes and user's rating for items.It adds the item attribute information into the decomposition matrix,and through a comprehensive analysis on the rating data set and attribute data set to get the user preference matrix and item attribute characteristic factor,it adds the user preference characteristic factor to the matrix decomposition.Experimental results on the datasets of Movielens and HetRec2011 show that in comparison with the classic matrix factorization collaborative filtering algorithm,the proposed algorithm can not only solve the problem in cold start of items to some extent,but also under the same condition reduce the root mean square error and the mean absolute error by 3％ and 4％,respectively.Especially on the more sparse HetRec2011 dataset,the impact of item attributes on user's rating behavior is more obvious,and this improved algorithm shows greater superiority in recommendation accuracy.