FedMLP4SR: Federated MLP-Based Sequential Recommendation System

Sequential Recommendation predicts users’ next possible item by modeling their historical interaction sequences. Transformer-based models can efficiently process chronological interaction sequences, but require quadratic model complexity. Traditional sequential recommendation system often adopts a centralized architecture, which easily leads to the leakage of user privacy. In our work, we propose a new sequential recommendation model based on Federated Learning called FedMLP4SR. On the one hand, MLP-based (Multi Layer Perceptrons) models have linear model complexity and can process sequential data without adding additional position embeddings. On the other hand, Federated Learning can meet the requirement of training the global model without collecting the private data from users. The MLP-based models have fewer parameters and less computational cost, making it highly adaptive to Federated Learning scenarios. Specifically, in our work clients utilize cross-sequence and cross-channel MLP blocks to capture correlational information in interaction sequences and complete the training of the local model; The server collects model parameters from clients to complete the updating of the global model. Besides, we use clients’ local graphs to extract high-order interactions between local items and fully consider item ratings. Finally, experiments validated the effectiveness and the feasibility of the proposed model over five representative baselines on two public datasets.