M3Net: A Multi-Metric Mixture of Experts Network Digital Twin with Graph Neural Networks

TL;DR

The paper tackles multi-metric network digital twins for 5G/6G environments by extending RouteNet-Fermi with a multi-metric mixture-of-experts (M3Net) GNN. It introduces a hierarchical approach to jitter and packet loss prediction and leverages sparsely gated MoE to handle diverse network scenarios, achieving improved delay accuracy (MAPE ~17.4%) and strong per-metric predictions (66.47% jitter, 78.7% packet-loss accuracy) on real-network data. A GPU-efficient implementation, including constant-batch flow concatenation and pre-computations, enables faster training and scalability. The work provides open-source tooling and demonstrates significant gains over baselines, underscoring the viability of accurate, scalable, multi-metric NDTs for network planning, slicing, and optimization.

Abstract

The rise of 5G/6G network technologies promises to enable applications like autonomous vehicles and virtual reality, resulting in a significant increase in connected devices and necessarily complicating network management. Even worse, these applications often have strict, yet heterogeneous, performance requirements across metrics like latency and reliability. Much recent work has thus focused on developing the ability to predict network performance. However, traditional methods for network modeling, like discrete event simulators and emulation, often fail to balance accuracy and scalability. Network Digital Twins (NDTs), augmented by machine learning, present a viable solution by creating virtual replicas of physical networks for real- time simulation and analysis. State-of-the-art models, however, fall short of full-fledged NDTs, as they often focus only on a single performance metric or simulated network data. We introduce M3Net, a Multi-Metric Mixture-of-experts (MoE) NDT that uses a graph neural network architecture to estimate multiple performance metrics from an expanded set of network state data in a range of scenarios. We show that M3Net significantly enhances the accuracy of flow delay predictions by reducing the MAPE (Mean Absolute Percentage Error) from 20.06% to 17.39%, while also achieving 66.47% and 78.7% accuracy on jitter and packets dropped for each flow

Figures (13)

• Figure 1: State-of-the-art model RouteNet-Fermi predicts delay well, but not jitter or packets dropped. Many flows experience no jitter or packet drops in our dataset.
• Figure 2: Histogram of delay, jitter and packets dropped in the training data. We see that all metrics are concentrated around zero, in particular the jitter and packets dropped. This concentration around zero makes it challenging to train predictors for these metrics, motivating Figure \ref{['fig:hierarchical_arch']}'s hierarchical approach.
• Figure 3: Hierarchical architecture for Jitter/Delay Estimation
• Figure 4: The MLP used in our hierarchical approach can predict whether a flow has zero jitter or packets dropped with high accuracy. We then directly predict 0 values for these flows.
• Figure 5: M3Net Architecture. We initialize the network with flow- and link-specific states, both of which are passed through separate MLPs. GRUs are then used for message passing through the links along the flow's path, and the output goes through the gating router to the readout MLPs. These MLPs' outputs are then combined into the final prediction.