Federated Recommendation with Integrated Dynamic Negative Sampling

The importance of recommendation system to solve the problem of information overload in the Internet era is self-evident. Implicit feedback is an important data source of recommendation system. By mining and analyzing implicit feedback data, recommendation system can provide better services. However, the wide application of recommendation system also makes the problem of user privacy protection increasingly prominent. The existing recommendation model often uses all the data of users for training, and does not take into account the subjective will of users, but users may not want to share all the data for security reasons. In addition, the proportion of negative samples in implicit feedback recommendation data is large, and random sampling is difficult for the recommendation model to learn useful information and make effective recommendations. To address the above shortcomings, this paper proposes a federal recommendation algorithm DNSFBPR, which integrates dynamic negative sampling strategy to select negative samples more effectively for the model, and designs a user control strategy on the client side to allow users to control whether private data is shared, while using homomorphic encryption to protect uploaded gradient information. Experimental results show that the proposed method can improve the recommendation performance while protecting user privacy.