Federated Learning with Adaptive Weighted Model Aggregation
Rebekah Wang · 2023
When using federated learning (FL) to predict user preferences, not all data samples are equally useful. For example, in video recommendation, recent watch history is more useful than older watch history. Thus, a new FL approach with adaptive weighted model aggregation (FLAMA) is proposed, where model aggregation weights are recalculated each FL training round based on the number of useful data samples used by each client and global model performance. Additionally, training datasets are compiled with a higher preference for useful data samples via prioritized data sample selection (PDSS). It is demonstrated that the proposed FLAMA achieves a higher accuracy in less FL training rounds compared to other schemes.