Throughput Prediction in Real-Time Communications: Spotlight on Traffic Extremes

Amidst the thriving advancement of networks, further catalyzed by the COVID-19 pandemic, we have witnessed a marked escalation in the worldwide adoption of Real-Time Communications (RTC) applications. In this context, there is a compelling necessity to cultivate intelligent and robust network infrastructures and technologies. Real-time throughput prediction emerges as a promising candidate for this purpose to foster network observability and provide preemptive functions, supporting advanced system management, e.g., bandwidth allocation and adaptive streaming. Nonetheless, contemporary solutions grapple with predicting extreme conditions in traffic throughput, notably peaks, valleys, and abrupt changes. To address the challenges, we propose a Transformer-based Deep Learning (DL) Neural Network (NN), leveraging solely packet-level information and adopting a multi-task learning paradigm, to predict short-term throughput, with an emphasis on critical values. In particular, our work is grounded in voluminous traffic traces procured from real video-teleconferencing sessions, and we formulate a time-series regression problem, comparing numerous technologies, from an adaptive filter to Machine Learning (ML) and DL approaches. Conclusively, our methodology exhibits superior efficacy, especially in forecasting traffic extremities.