DUPGT-CDR: Deep User Preference Gating Transfer for Cross-Domain Recommendation

Cross-domain recommendation (CDR) seeks to enhance recommendation accuracy in a target domain by utilizing user–item interactions from a source domain, especially in cold-start scenarios. In conventional CDR approaches, user-specific bridge functions generated using only the user’s high-rated interaction history in the source domain are used to map user embeddings from the source domain to the target domain. As a result, these methods cannot effectively leverage both positive and negative feedback. To address this limitation, we propose DUPGT-CDR, a new framework that encodes high (ratings 4–5) and low (ratings 1–3) feedback separately, and then adaptively fuses these signals using a gating network. This gating mechanism enables fine-grained control over transferring high and low feedback. Experiments on three real-world domain pairs with varying overlap ratios demonstrate that DUPGT-CDR consistently outperforms state-of-the-art methods (EMCDR, PTUPCDR, DPTUPCDR, MIMNET), achieving up to17.9%lower MAE and20.9%lower RMSE, thereby confirming its robustness across different data partition settings. The code is publicly available at https://github.com/isdlyusukeshimizu/DUPGT-CDR.