FedSarah: A Novel Low-Latency Federated Learning Algorithm for Consumer-Centric Personalized Recommendation Systems

Data heterogeneity, insufficient scalability, and data privacy protection are the technological challenges of personalized recommendations. This study proposes a new federated learning algorithm (FedSarah) to address low scalability caused by data heterogeneity and uneven computing power in consumer-centric personalized recommendation systems while protecting data privacy of consumers. The algorithm updates the stochastic gradient estimates using a recursive framework on consumer clients. The outer loop calculates the entire gradient for updating global model, and the inner loop calculates the stochastic gradient based on the accumulated stochastic information for updating local models. To increase the stability of convergence, the inner loop modifies intrinsic parameters to change the number of training rounds and the direction of model update on consumer clients. The detailed mathematical analysis and experiments demonstrate that FedSarah has good convergence. In addition, it's shown that the algorithm can achieve a performance improvement of nearly 5% in terms of accuracy compared to the traditional FedAvg and FedProx algorithms under the condition of heterogeneous data. Furthermore, under the condition of effective privacy protection on consumers' data, the new algorithm can significantly lessen the impact of data heterogeneity on the real-time service of consumer-centric personalized recommendation systems with low communication latency. The code is available at https://github.com/DashingJ-82/FedSarah.git.