Adaptive windowing-based concept drift detection and adaptation framework for human-to-machine applications over future communication networks

Abstract

Human-to-machine (H2M) applications in future networks have strict low-latency transmission requirements. Thanks to the assistance of machine learning (ML) in future communication networks, ML-enhanced dynamic bandwidth allocation (DBA) schemes have been proposed to effectively reduce uplink latency in H2M applications. Existing methods generally assume that the H2M application traffic stream is stationary, thereby primarily designing an ML model based on a specific H2M application and fixed traffic load. However, future communication networks are expected to support dynamic and heterogeneous H2M applications. As such, incoming H2M data traffic from networks will change over time as different H2M applications have distinct traffic distributions, causing the concept drift in H2M applications. Meanwhile, another challenge in detecting the concept drift of H2M applications is detecting traffic distribution change in dynamic network environments among similar H2M application scenarios, which leads to an incremental drift of H2M application traffic. To tackle the above challenges, we propose an adaptive windowing-based concept drift detection and adaptation (ADA) framework to support H2M applications in dynamic and heterogeneous networks. Unlike existing solutions that mainly use fixed sliding windows, the proposed ADA dynamically changes the sliding window size based on the drift detection results. Hoeffding’s inequality-based drift detection algorithm is employed in ADA to effectively detect incremental H2M application traffic drift in a dynamic network. Comprehensive simulation investigations show that ADA can enhance DBA performance in terms of uplink latency reduction and rapidly responding and adapting to the concept drift in changing H2M applications and traffic load scenarios of 96.51% drift detection efficiency improvement, and over 50% packet delay reduction in model adaptation compared to frameworks are considered.