Enhancing Cross-domain Recommendation through Preference Structure Information Sharing

Abstract

Cross-domain recommendation can alleviate data sparsity problems by leveraging data from multiple domains. Hence it is becoming an emerging research topic. Existing approaches based on latent factor models usually attempt to utilize cross-domain information in the form of inner product or Euclidean distance. This kind of form only enables the model to incorporate non-structure similarity relations between entities. However, the studies in the literature of consumer behavior have disclosed that consumers make decisions following a structural characteristic. Inspired by this, we propose a new cross-domain recommendation model, named PSRec, which learns and shares users' inherent preference structure information during their decision making in relation to their product choices in different domains. In PSRec, the latent factors of items of different domains are mapped to a common space so that the factorized user latent vectors have similar preference structures. Moreover, users' preference structure information is transferred from an auxiliary domain to a target domain as a constraint to a non-negative matrix factorization algorithm. Extensive experiments on two real-world datasets demonstrate the effectiveness and competitive performance of PSRec compared to several state-of-the-art solutions.