Abstract:

The cold-start problem of recommendation systems which affects recommendation quality, service experience, privacy, and security, has become one of the most challenging research hotspots in the field. Thus, our study proposed an integrated federated learning framework for conditional variational autoencoders (CVAE), termed as FedCVAE. Individual CVAE models were trained on each client’s local data to generate embeddings of users, items, and user interaction sequences. These embeddings were used as inputs for the essential recommendation model. The model’s global parameters were aggregated and updated at the server-side, which was subsequently disseminated back to the client-side to support local CVAE models in updating hyperparameters. While the model improves its accuracy in handling sparse data, it also enhances its ability to preserve privacy, thus effectively mitigating the cold-start problem. The experimental results indicate that in three typical cold-start scenarios, the model presented in this paper outperformed mainstream recommendation algorithms. The mean absolute error metric reduces by approximately 0.8% ~ 5.5%, and the Hit@5 metric improves by approximately 1.2% ~ 5.7%. The model demonstrated superior performance, delivering high-quality recommendation services that balance personalized experiences and enhanced privacy protection.

Key words: recommendation systems, cold-start, federated learning, meta-learning, variational autoencoder