Toward Robust End-to-End Delay Prediction: A GNN Approach With Routing-Aware Attention and Masked Subgraph Sampling

The end-to-end delay prediction is critical for intelligent network management, particularly in latency-sensitive and dynamic environments. While recent deep learning (DL) models have shown promising results, their reliance on sequential encoding of routing paths limits generalization to unseen routing schemes. In this article, we propose a robust graph neural network (GNN)-based delay prediction model that overcomes this limitation by introducing a global routing representation and a routing-aware attention mechanism. The model queries flow-relevant features from a unified topology-routing map without relying on routing sequences. In addition, a mask-based subgraph sampling strategy enables the model to infer global routing correlations from partial flow interactions, further enhancing its adaptability. Extensive experiments are conducted on four public datasets-TnCwD, NSFNET, GBN, and GEANT2. The results demonstrate that our model not only outperforms existing methods in prediction accuracy but also exhibits strong generalization across diverse routing configurations. Future work will focus on further validating the model under more complex and dynamic routing scenarios.