TransMUSE: Transferable Traffic Prediction in MUlti-Service Edge Networks.

Area of Science:

• Computer Science
• Network Engineering
• Artificial Intelligence

Background:

• The COVID-19 pandemic shifted workforces to remote models, increasing demands on broadband network management.
• Accurate network traffic prediction is vital for maintaining reliable connectivity, especially at the network edge.
• Existing neural network models struggle with traffic prediction across different geographic regions due to data variability.

Purpose of the Study:

• To develop a novel deep learning framework for transferable network traffic prediction in multi-service edge networks.
• To address the limitations of bespoke model training, including high measurement overhead and computational costs.
• To enable accurate, fine-grained demand forecasts for edge services provisioning.

Main Methods:

• Proposed TransMUSE (Transferable Traffic Prediction in MUlti-Service Edge Networks) framework.
• Clustering similar services and grouping edge nodes into cohorts based on traffic feature similarity.
• Utilized a Transformer-based Multi-service Traffic Prediction Network (TMTPN) for direct transfer within cohorts.

Main Results:

• TransMUSE demonstrated imperceptible performance degradation in Mean Absolute Error (MAE) compared to individually trained models.
• The TMTPN architecture achieved up to 43.21% lower MAE than state-of-the-art methods.
• Successfully reduced measurement overhead while providing accurate traffic forecasts.

Conclusions:

• TransMUSE offers an effective and scalable solution for network traffic prediction in multi-service edge environments.
• The framework enables model transferability across similar network traffic cohorts, minimizing customization needs.
• This work pioneers the joint use of model transfer and multi-service prediction for efficient edge network management.