An Improved NCF Model in Federated Recommendation Systems

Due to the communication of federated learning and the privacy protection in neural collaborative filtering recommendation. An Improved NCF model with federated learning is proposed by merging weight setting based on participation, pretraining, local parameters and Differential Privacy. This improved NCF model in federated recommendation include three sub-models: generalized matrix factorization, multilayer perceptron and neural matrix factorization. First, a weighting algorithm based on participation is proposed to improve the efficiency of federated learning. Secondly, the local parameters of the aggregate model are separated to protect the user embedding layer. Finally, the Laplacian noise in differential privacy is adding to protect the uploaded parameters during the federated learning training. A series of comparative experiments are conducted to validate those federation recommendation sub-models proposed. The results shown those models can achieve a compromise between recommendation effect and privacy protection.