In today's era, users have increasingly high expectations regarding the performance and efficiency of communication networks. Network operators aspire to achieve efficient network planning, operation, and optimization through Digital Twin Networks (DTN). The effectiveness of DTN heavily relies on the network model, with graph neural networks (GNN) playing a crucial role in network modeling. However, existing network modeling methods still lack a comprehensive understanding of communication networks. In this paper, we propose DWNet (Deeper and Wider Networks), a heterogeneous graph neural network modeling method based on data-driven approaches that aims to address end-to-end latency and jitter prediction in network models. This method stands out due to two distinctive features: firstly, it introduces deeper levels of state participation in the message passing process; secondly, it extensively integrates relevant features during the feature fusion process. Through experimental validation and evaluation, our model achieves higher prediction accuracy compared to previous research achievements, particularly when dealing with unseen network topologies during model training. Our model not only provides more accurate predictions but also demonstrates stronger generalization capabilities across diverse topological structures.

翻译：当今时代，用户对通信网络的性能和效率要求日益提高。网络运营商希望通过数字孪生网络实现高效的网络规划、运维与优化。数字孪生网络的有效性高度依赖于网络模型，其中图神经网络在网络建模中扮演着关键角色。然而，现有的网络建模方法仍缺乏对通信网络的全面理解。本文提出DWNet（更深更宽网络），一种基于数据驱动的异构图神经网络建模方法，旨在解决网络模型中的端到端时延与抖动预测问题。该方法因两大特点而脱颖而出：其一，在消息传递过程中引入更深层次的状态参与；其二，在特征融合过程中广泛整合相关特征。通过实验验证与评估，我们的模型相较于先前研究成果实现了更高的预测精度，特别是在处理模型训练过程中未见过的网络拓扑时表现尤为突出。我们的模型不仅提供更准确的预测，而且在不同拓扑结构间展现出更强的泛化能力。