Abstract: Matrix Factorization algorithm is widely used in recommendation systems.However,its performance is often affected by data sparsity and high dimensionality of the data,and less consideration is given to the content information of items.In view of the above problems,a Bipartite Network Representation learning regularized Matrix Factorization recommendation algorithm (BiNRMF) was proposed.Two bipartite networks were constructed using the rating information and tag information of the item respectively.Low-dimensional vectors of users and items were obtained by Bipartite Network Embedding (BiNE) algorithm,which could be applied to calculate the similarity between users and between items.Traditional matrix factorization model was regularized by these similarities.Experimental results on GoodBooks and MovieLens datasets demonstrated a significant improvement in prediction accuracy compared to the classical recommendation algorithm.

Key words: recommendation systems, collaborative filtering, bipartite network, network representation learning, matrix factorization

摘要： 矩阵分解算法广泛应用于推荐系统。然而,其性能往往受到数据稀疏性和数据高维度的影响,且较少考虑项目的内容信息。针对上述问题,提出一种联合二部网络表示学习的矩阵分解推荐算法(BiNRMF)。首先,利用评分信息和项目的标签信息构建两个二部网络;然后,通过二部网络的表示学习算法得到用户和项目的低维向量表示,用以计算用户之间和项目之间的相似性;最后,改进传统矩阵分解模型,融入低维向量空间中用户的相似关系和项目的相似关系。在GoodBooks和MovieLens数据集上的实验结果表明,与经典的推荐算法相比,联合二部网络表示学习的矩阵分解推荐算法的预测精度有显著提升。

关键词: 推荐系统, 协同过滤, 二部网络, 网络表示学习, 矩阵分解