September 7, 2024

Recently, the acceptance results for the ACM International Conference on Information and Knowledge Management (CIKM 2024) were announced. The paper titledEFVAE: Efficient Federated Variational AutoEncoder for Collaborative Filtering, authored bySSEPhD student Zhang Lu under the guidance of Professor Li Guohui, has been accepted.

This paper focuses on addressing the issue of excessive training communication costs inFederated Recommender Systems. In recommendersystems, federated learning has become an important method for tackling privacyissues. However, existing federated variational autoencoders (FedVAE) face high communication costs during the training process due to their model parameters being related to the number of items, whichmakes them lesspractical.

To address these challenges, we propose an Efficient Federated Variational AutoEncoder for collaborative filtering, EFVAE, which core is the Federated Collaborative Importance Sampling (FCIS) method. FCISsignificantly reduces communication costs during model training. This method dynamically approximates the decoding distribution through a collaborative sampling mechanism between the client and the server, thereby substantially lowering communication costs while maintaining recommendation performance. Extensive experimentsshow that EFVAE not only effectively reduces communicationcostsbut also demonstrates excellent recommendation performance, particularly exhibiting clear advantages in sparse data scenarios. Additionally, FCIS can also be applied to other autoencoder-based recommendation systems, further showcasing its applicability and effectiveness in various contexts.

CIKM(ACM International Conference on Information and Knowledge Management)is a B-level conference recommended by the China Computer Federation (CCF). CIKMisalsoone of the important academic conferences in the fields of information retrieval and data mining.