Graph Neural Network-based Delay Prediction Model Enhanced by Network Calculus

Network modeling is critical to network management. Existing network modeling approaches still have room for improvement in performance prediction of important performance metrics, such as latency. Graph Neural Network (GNN) has proven their effectiveness for network performance prediction. However, the shortcomings of GNN in data interpretability limit their performance. To address this problem, we introduced network calculus to enhance the interpretation and learning of GNN for data. We proposed a modified GNN model NetCTRT, which understands the information between topology, routing and traffic and interacts with the Network Calculus (NC)-based delay bound so that the end-to-end delay can be accurately predicted. Experimental results show that NetCTRT can significantly improve the prediction accuracy.