Uncertainty quantification and consideration in ML-aided traffic-driven service provisioning

The network traffic prediction problem has been extensively studied in the literature through the application
of statistical linear models and more recently through the application of machine learning (ML). In fact, ML
has proven its capabilities on accurately modeling the non-linear nature of network traffic, outperforming
conventional statistical linear models. Without doubt, model accuracy constitutes an important evaluation
metric. In a network environment, however, where uncertainty may lead to erroneous service provisioning
decisions (e.g., violations of the quality-of-service (QoS) requirements), it does not provide any information
of how much the traffic predictions can be trusted. Hence, in this work the focus is on addressing traffic
prediction uncertainty by leveraging the capabilities of Monte Carlo (MC) dropout inference. The proposed
framework is compared with a margin-based technique, traditionally used to compensate for traffic prediction
uncertainty, demonstrating that the MC dropout inference framework results in significant spectrum savings,
as opposed to the baseline, myopic, margin-based scheme. Even though a small penalty is observed on the
unpredictable traffic, this is successfully handled on-line with a reduced operational overhead compared to the
case where prediction uncertainty is completely ignored. Importantly, it is shown that, unlike the margin-based
framework, the MC dropout inference framework can be used for the provisioning of services with diverse QoS
requirements.