Cross-Domain Recommendation With Personalized Rating Pattern Compatibility as Transfer Rate Between Domains

Cross-Domain Recommendation Systems (CDRs) address the data sparsity issue by transferring useful information from one domain to another. Despite advancements, these systems often face two main limitations. First, existing models transfer information without considering the magnitude of differences between domains, which can result in inaccurate information transfer when the characteristics of the domains diverge. Secondly, they prioritize domains based solely on user interaction counts, which biases the model toward high-frequency domains (e.g., Food) over low-frequency ones (e.g., Cars). To overcome these limitations, we propose a CDR that utilizes adaptive domain information transformation based on domain compatibility while prioritizing each domain’s information according to its relative domain density. Our proposed method compares the rating pattern of the target user across domains to calculate the compatibility level, which is used in the transfer function. This process prevents the leakage of irrelevant information by allowing the model to apply a lower transfer rate when the characteristics of the domains differ and a higher transfer rate when they are similar. To alleviate the bias caused by relying solely on the number of interactions, the proposed method prioritizes dense domains based on relative density. Specifically, relative density is computed by considering the ratio of the number of items in each domain to the number of interactions the target user has within that domain. In the experiments, the proposed method consistently outperformed all baselines across three benchmark datasets.