Federated Contrastive Learning for Cross-Domain Recommendation

Abstract

Conventional cross-domain recommendation models, which require centrally collecting varieties of original data from users, usually meet a challenge that users are reluctant to provide their real feedbacks because of privacy concerns. Thus, federated learning is incorporated into cross-domain recommendation, since it aggregates parameters of local models trained on user sides to train a global recommendation model, instead of centralized data collection. However, the deviations between the global model and local ones, which are caused by users’ data with non-independent and identical distributions, significantly challenge existing federated learning-based models in terms of alleviating data sparsity and cold-start problems. This article proposes a novel end-to-end federated contrastive learning-based model towards cross-domain recommendation, namely ${{Fed-CLR}}$. It first uses an inference model to characterize interaction distributions of users in source domain(s), then reconstructs historical interactions of users in target domain(s) with a generative model, and finally performs federated contrastive learning at model level (including inner-model and inter-model) to help reduce deviations between the global model and local ones. Particularly, a constraint mechanism, namely ${{Con-Mec}}$, is proposed to achieve consistency reinforcement from the aspect of inner- and inter-models. The experimental results on three real-world datasets not only show that ${{Fed-CLR}}$ outperforms the state-of-the-art peers, but also demonstrate that ${{Fed-CLR}}$ achieves a faster convergence speed than classical federated learning-based models.