基于迁移的联合矩阵分解的协同过滤算法

Collaborative filtering recommendation based on transfer learning and joint matrix decomposition

摘要:

早期的协同过滤算法利用矩阵分解来解决数据稀疏问题，但是严重的稀疏问题导致矩阵分解的性能很难满足应用的需求.随后，迁移学习被引入到协同过滤的研究中，它主要利用辅助域和目标域的公共用户的各种信息来解决目标域的数据稀疏问题.虽然通过引入辅助域的信息能够帮助目标域获取更多的知识，但是在公共用户包含的公共商品项目少的情况下，只利用公共用户的浅层特征来度量用户的相似性，不能很好地捕捉用户的潜在特征，相似性度量效果不好.为此，本文提出了一种基于迁移的联合矩阵分解协同过滤模型，以公共用户为锚，将两个领域的用户和商品映射到一个潜在的语义空间.模型通过对两个领域的用户商品评分矩阵在以公共用户信息作为约束项的情况下，进行联合矩阵分解，在实际基准数据集上的实验结果表明，本文所提出的方法明显优于现有基于相似度计算的迁移学习方法，也证明了模型的有效性.

Abstract:

Matrix decomposition was used in the early collaborative filtering algorithms in order to solve the problem of data sparsity. But it performed poorly in handling serious sparsity problem and cannot meet the application requirements. Then, transfer learning was introduced into collaboration filtering to deal with the data sparsity in the target domain by utilizing common users’ information in the auxiliary and target domains.Although the introduced auxiliary information would prompt knowledge acquisition in the target domain, these methods only use shallow features to measure the users’ similarity. As a result, these methods could not capture the potential features when the users have only a few common items and would result in poor performance in similarity measurement. In order to address these problems, this paper proposes a collaborative filtering recommendation method based on transfer learning and joint matrix decomposition. In this method, the information of common users and items in the two domains is mapped into a potential semantic space with the information of users as anchors; the useritem joint rating matrix of two domains is decomposed with the user information as the constrain. The experiment was performed to validate the proposed method and the method showed superior performance over the stateoftheart migration learning methods based on similarity calculation on benchmark data set, proving its effectiveness.