Oh-FedRec: One-Shot and Heterogeneous Vertical Federated Recommendation System

Abstract

Federated learning has a wide range of applications in recommendation systems, but most federated recommendation systems can only achieve federated communication between users and servers. Only a few are vertical federated recommendation systems, achieving federated server communication. In addition, the current federated recommendation frameworks require that each participant have the same model. This condition is very harsh for all participants involved. What is more, the current federated recommendation frameworks require multiple rounds of communication, which consumes many communication resources and much time. To solve these problems, we propose a one-shot and heterogeneous vertical federated recommendation framework called Oh-FedRec. This framework needs only one round of communication, dramatically reducing the consumption of communication resources. Additionally, it no longer requires participants to have consistent models, significantly reducing constraints on the various participants. We tested our framework on Tmall and Jingdong datasets, and the test results proved the superiority of Oh-FedRec.