基于主题模型的矩阵分解推荐算法

Matrix factorization recommendation based on topic model

针对协同过滤算法存在的数据稀疏和忽视用户喜好多主题的问题,提出了基于主题模型的矩阵分解推荐算法,将标签、主题模型融合到了矩阵分解模型当中。该方法首先根据物品的标签提取物品的主题特征,用物品主题特征向量表达该物品,然后通过相似度计算方法得到每个物品的最近邻,最后用基于最近邻的正则化项来改进矩阵分解模型。在实验分析中,选择了不同的主题数进行比较,并且在潜在因子数不同的情况下,对比了该算法和潜在因子模型、正则化奇异值分解推荐算法。实验结果表明,改进算法能够降低预测评分的均方根误差,提高评分预测的准确度。

In order to solve the problems of data sparseness and neglecting the multi-themed of user preferences existed in collaborative filtering algorithm,a kind of matrix factorization recommendation algorithm based on the topic model was proposed. This method integrated label and topic model into matrix factorization model. First,tags of item were classified as topics. Then the similarity among items was calculated according to the topics of items. Last the nearest neighbors of each movie was found using particular similarity calculation method. The information of nearest neighbors was used as regularization term to improve the matrix factorization model. Different number of topics was used to compare this method with other recommendation algorithms using different latent factor numbers. The experimental result shows that the proposed method is able to reduce the root mean squared error between predicted ratings and true ones,and improve the accuracy of predicted rating.