FLAIR: Federated Learning via Adaptive Influence and Resource Coordination

Adaptive video streaming (also known as adaptive bitrate or ABR streaming) must intelligently balance competing Quality of Experience (QoE) factors-video quality, rebuffering, and smooth playback-under widely varying network conditions. Traditional learning-based ABR algorithms struggle to generalize across users' diverse network environments and QoE preferences [21] [22]. Generic federated learning (FL) approaches, on the other hand, often suffer in this setting due to client heterogeneity (network speeds, device capabilities) and multi-faceted QoE objectives that standard FL algorithms do not explicitly optimize. In this work, we propose FLAIR, a Federated Learning-based framework that employs Adaptive Influence and Resource Coordination for video streaming. FLAIR holistically addresses the challenges of training a global ABR policy model across many users without collecting their data. It introduces: (1) QoE-weighted model aggregation to prioritize updates from clients in critical streaming conditions, (2) influence-guided client selection via an influence matrix that identifies clients most impactful to improving global QoE, and (3) entropy-based scheduling of workload to dynamically include clients with novel or changing network conditions. We integrate a new Influence-Entropy-Diversity (IED) metric directly into FLAIR's optimization objective and convergence analysis, proving it can accelerate convergence by reducing gradient variance. Extensive experiments on both real- world streaming traces and large-scale synthetic scenarios demonstrate that FLAIR significantly improves average QoE (e.g., higher video quality and fewer stalls) compared to conventional federated training and state-of-the-art methods. Notably, FLAIR's global model adapts swiftly to network shifts (such as WiFi-to-cellular handoffs) and scales to thousands of clients, indicating its practical deployability for internet-wide adaptive streaming.