Graph Collaborative Filtering and Data Augmentation Strategies in Dual-Target CDR

Current dual-target recommendation methods focus on efficient feature fusion but neglect the inherent noise issues in the domains. However, noise negatively affects the fusion of domains. To tackle this issue, we introduce an improvement and noise reduction strategy named DA-DCDR(Data Augmentation-Dual Target Cross Domain Recommendation), for domain fusion. By refining and reducing noise in each domain’s subgraphs, we not only enhance the accuracy of interaction data but also ensure consistency in data scales. To establish associations between distinct domains, we implement a graph co-training strategy. Key procedures of DA-DCDR include interaction refinement and noise reduction, domain fusion, and correlation expansion. We use graph encoders to acquire user/item embeddings for both domains before domain fusion, followed by enhancement and noise reduction in interactions via top-k sampling and re-prediction. Additionally, we amplify user-user and item-item correlation elements after the domain fusion. Experimental results validate the noteworthy performance enhancement of our proposed strategy in the dual-target recommendation, mitigating the noise effects and boosting the accuracy of the dual-target recommendation system.