Federated recommender systems based on deep learning: The experimental comparisons of deep learning algorithms and federated learning aggregation strategies

Due to the requirements of privacy protection and data asset ownership, recommender systems (RSs) based on centralized training process may be infeasible in some practical situations. In these situations, federated recommender systems (FedRec) based on deep learning are regarded as the substitutes with extensive application backgrounds and enormous potential business values. However, which deep learning algorithm and federated learning aggregation strategy perform best in the training of FedRec is still an unsolved question. To answer this question, experiments are conducted for comparing the performances of four popular deep learning algorithms and three federated learning aggregation strategies in FedRec. The experiments are conducted on 8 public datasets to obtain 96 FedRec precisions (4 deep learning algorithms x 3 federated learning aggregation strategies x 8 datasets). Then, the Wilcoxon test is used to verify whether there is a significant difference between the FedRec combinations and to derive the best combination. The result shows that, in terms of recommendation precision, the best FedRec is the combination of neural FM and federated averaging. In terms of execution iterations, the best FedRec is the combination of deep crossing and attentive federated aggregation. The obtained results would be valuable for further promoting the development of RSs under the situations that a centralized training process is infeasible.