User-centric User Selection for Federated Learning in Visible-Light Networks

In order to increase the number of users involved in federated learning (FL) and further to reduce the training loss, a joint user-centric (U C) cluster formation and user selection problem in visible-light networks is studied in this paper. The original problem is non-convex and a heuristic distance-based algorithm is proposed by taking the training data size of each user into account, while satisfying the delay requirement. Based on a handwritten digit identification task, our simulation results demonstrate that the UC scheme is superior than the traditional network-centric (NC) design in terms of the total number of users participating in FL as well as the identification accuracy in all scenarios considered.