Data Sampling In Federated Learning: Principles, Features And Taxonomy

Abstract

Federated learning collects data from various devices, analyzes it locally, aggregates it, and then finds meaningful insights from it. Data sampling works the same way by dividing the larger data set into smaller parts and applying computation to those data sets, which reduces the time taken to do the work. Data sampling in federated learning aims to find the ideal mixture of selecting data sets for training purposes to improve training accuracy while staying within the maximum capability of the device and network. In this article, we present an overview and analysis of recent data sampling techniques for federated learning. The list includes sampling approaches suitable for federated learning environments such as clustering, dynamic sampling, adaptive sampling, probabilistic sampling, and many more. The feature analysis is comprised of a description of the procedure, the criteria, and other relevant parameters for sampling. The efficiency of the sampling technique is analyzed via comparison of claimed accuracy and convergence rate with respect to the used dataset.