Research on improving a graph neural network model forcomputer network simulation

Abstract

Modern computer networks face increasing challenges due to the growing complexity of their structure, dynamic traffic fluctuations, and the need to maintain high performance. Traditional approaches to network modeling often fail to accurately predict parameters such as latency or packet loss, as they have limited capacity to capture the specific characteristics of individual network elements. This highlights the relevance of developing novel methods that can adapt to real operating conditions and ensure efficient resource management. The aim of this study is to enhance network modeling methods by developing a model that incorporates the individual properties of network elements to improve the accuracy of parameter prediction and to optimize routing processes. The research objectives include the analysis of current modeling approaches, the design of an improved model based on machine learning techniques, the refinement of training algorithms, and the execution of experiments to evaluate the model’s effectiveness. Machine learning methods were applied in the implementation, with particular emphasis on a graph neural network, which enables the modeling of complex interdependencies among network elements. The proposed model integrates node-specific characteristics into the data processing pipeline, thereby ensuring adaptability to heterogeneous conditions. Experiments were conducted on multiple datasets representing real-world network topologies; with prediction accuracy assessed using several evaluation metrics. The results demonstrate that the proposed model provides higher accuracy in predicting network parameters compared to baseline approaches, exhibiting the ability to generalize to unseen topologies. The scientific novelty of the work lies in the incorporation of element-level characteristics into the modeling process, allowing for a more precise reflection of real-world conditions. The practical significance is manifested in the potential application of the model in network management systems for routing optimization and infrastructure cost reduction. The findings open new prospects for further development of modeling and management methods in modern networked systems.