In-network Predictive Analytics in Edge Computing

Abstract

Edge-centric predictive analytics methodologies use real-time model caching to significantly reduce the communication overhead. We investigate an approach of using different regression techniques at the edge as caching models. Our methodology reports on an edge-centric mechanism to automatically decide when to update the parameters of the cached models to a central location (data center). Through experimentation, we showcase the trade off between accuracy and communication overhead and conclude that for all the experimented regression models, a lower percentage of the cached models should be sent to the data center to significantly decrease the communication overhead.