Federated cross-domain recommendation system based on bias eliminator and personalized extractor

Yicheng Di, Hongjian Shi, Qi Wang, Shunyuan Jia, Jiayu Bao, Yuan Liu

Abstract: The rapid development of big data is driven by advancements in Internet of Things technology. The cross-domain recommendation system is a highly successful approach for obtaining user-interested items from massive data. However, implementing a cross-domain recommendation system on distributed IoT devices faces three challenges. On the one hand, item embeddings used in existing cross-domain recommendation models often lead to an unavoidable popularity bias. On the other hand, to convey user preference information, most existing methods utilize common interest channels. However, it stands to reason that the interest channels for different users should be distinct, as each person has their own unique preferences. Furthermore, collecting raw data from distributed IoT devices may lead to user privacy concerns. Given these challenges, we propose a federated cross-domain recommendation system based on bias eliminator and personalized extractor (FedBP) in this paper to achieve precise recommendations in cold-start scenarios. Firstly, we employ a bias eliminator to unfold all embedding directions during training, ensuring that each direction captures only specific features while maintaining neutrality in popularity. Secondly, personalized extractor is utilized to individualize the distinct preference information of each user from the source domain to the target domain. Then, we utilize a federated framework to collaboratively train the cross-domain recommendation system model, where local differential privacy is employed to ensure data privacy. Experimental results on public benchmarks show that FedBP consistently outperforms baseline models across three cold-start cross-domain recommendation scenarios, with improvements of at least 3.02%, 5.08%, and 3.08%.