FedOCD: A One-Shot Federated Framework for Heterogeneous Cross-Domain Recommendation

Cross-domain recommendation (CDR) has been widely applied to address the issue of data sparsity by transferring information from the source domain to the target domain. Since direct information transfer leaks data privacy, existing work combines federated learning with CDR. However, these approaches require that the model of each domain be homogeneity and have high communication overhead, which limits the applicability of federated learning to some extent. In this paper, we propose FedOCD, a one-shot federated cross-domain ensemble Learning framework. FedOCD consists of two main components: source domain modeling aims to learn user embeddings from the source domain data while ensuring privacy through Local Differential Privacy (LDP), and target domain ensemble aims to synthesize the user features from the source domain to improve the recommendation performance. Experimental results on benchmark datasets demonstrate that FedOCD has made significant improvements.