User and item spaces transfer from additional domains for cross-domain recommender systems

Data sparsity is one of the common problems to provide recommendations. Transfer learning is effective as a potential solution to build more effective recommendation models, which is called cross-domain recommender systems. It transfers some useful knowledge from additional data sources to alleviate the said problem to desirable domains. The types of knowledge for transferring can be user side, item side, or a combination of both sides. Several researchers have provided solutions based on types of knowledge and transfer strategies with their limitations. This paper proposes a method, namely a User and Item Spaces Transfer (UIST), that transfers the learned knowledge from two independent source domains. The agenda of transfer learning in a recommender system is to improve the accuracy performance of the target domain. The first domain is used for extracting user side knowledge, and the second domain is used for extracting item side knowledge. The novelty of our proposed work over existing methods is that UIST can capture knowledge independent from the domain and control the dominating behavior of source domain knowledge. Therefore, our work is capable of transferring both sets of information more precisely. The proposed work has been verified and claimed by the experimental results on three real-world datasets.