Novel Concept Drift Detection and Adaptation (CDDA) Framework for Human-to-Machine (H2M) Applications over Future Communication Networks

Machine learning (ML)-enhanced future communication networks are able to advance human-to-machine (H2M) applications by intelligent bandwidth prediction techniques to achieve bandwidth pre-allocation. Existing methods of H2M bandwidth prediction typically assume the stationary data stream over time. However, in the near future, communication networks are expected to support dynamic and heterogeneous applications. Since different H2M applications will exhibit different traffic distributions and loads across the day, an ML model learned on a specific H2M application at a particular network load will, therefore, be unable to adapt to changing applications and network loads. This will give rise to the phenomenon known as concept drift. This paper addresses concept drift in dynamic and heterogeneous networks supporting H2M applications by proposing a novel framework, the concept drift detection and adaptation (CDDA) framework, to respond and adapt to the concept drift rapidly. CDDA learns the traffic characteristics of H2M applications and combines offline and online learning processes to enhance H2M traffic prediction and improve band-width prediction performance. Results from our investigation using experimental traffic from H2M applications over a 10Gb/s passive optical network simulator show that CDDA can more rapidly respond to concept drift and better predict the bandwidth of changing H2M applications and network load.