NetMOS: Topology-aware VoIP MOS prediction via attention-recurrent GNNs

Author Identifier (ORCID)

Wei Ni’s ORCID record ORCID Logo

Abstract

The Mean Opinion Score (MOS) is a standard metric for assessing the Quality of Experience (QoE) in Voice over IP (VoIP) applications. Accurate prediction of how network conditions influence MOS is critical for network planning, operation, and optimization. This requires modeling traffic flows with application-level granularity, which significantly increases both the dimensionality and structural complexity of the learning task. The ability to achieve efficient, robust, and generalizable data-driven learning in the presence of such complexity depends critically on the careful design of model architectures. This paper presents NetMOS, a Graph Neural Network (GNN) architecture specifically crafted to model IP networks and predict VoIP MOS scores. NetMOS models IP networks as heterogeneous graphs and designs a two-stage Message Passing Neural Network (MPNN) to capture both permutation invariant and sequential dependencies in traffic flow and network interactions. It uses a Gated Recurrent Unit (GRU) layer to model the ordered influence of links along a traffic path and introduces a customized attention layer with Sigmoid activations to model the cumulative effects of multiple flows on the links. Simulations demonstrate that NetMOS consistently outperforms conventional GNN-based baselines across diverse network topologies in Mean Absolute Error (MAE), R2 score, Pearson correlation, and Spearman correlation. NetMOS generalizes effectively beyond the training topology, maintaining high prediction accuracy on unseen network topologies and varying network activity durations without retraining. NetMOS also provides MOS predictions 44× – 170× faster than packet-level simulations.

Keywords

graph neural networks, mean opinion score, network modeling, quality of experience, VoIP

Document Type

Journal Article

Date of Publication

1-1-2026

Volume

23

Publication Title

IEEE Transactions on Network and Service Management

Publisher

IEEE

School

School of Engineering

Copyright

subscription content

Content Type

Metadata only

First Page

5280

Last Page

5292

Recommended Citation

Ranaweera, S., He, Y., Jayawickrama, B., Wang, X., Liu, R. P., & Ni, W. (2026). NetMOS: Topology-aware VoIP MOS prediction via attention-recurrent GNNs. IEEE Transactions on Network and Service Management, 23, 5280–5292. https://doi.org/10.1109/TNSM.2026.3701762

Share

 
COinS
 

Link to publisher version (DOI)

10.1109/TNSM.2026.3701762