Network Delay Prediction Based on Graph Neural Network and Attention Mechanism

Network modeling serves as a powerful tool for network management and optimization, enabling network operators to more accurately understand the network state by predicting critical performance indicators, such as delay. Existing network modeling methodologies suffer from several inherent limitations, including excessive reliance on idealized assumptions, prohibitive computational complexity in simulation processes, and lack of effective capture of the relationship between network elements. To solve these problems, a delay prediction method combining the graph neural network and the self-attention mechanism is proposed in this paper. The proposed architecture achieves dual optimization through complementary technical enhancements: The self-attention module compensates for the deficiencies of conventional recurrent neural networks in modeling long-range dependencies during the message-passing phase, thereby enhancing global context awareness in graph neural networks; simultaneously, the graph neural network enhances the ability of the self-attention mechanism to process non-Euclidean structured data, promotes the integration of global information and local features, and achieves topological generalization. Experimental results on real-world datasets demonstrate that the proposed method yields more accurate delay predictions than the baseline methods when faced with unknown topologies.