✨ TL;DR
This paper proposes a Graph Neural Network model to predict network flow-level traffic (NetFlow) by representing network traffic as heterogeneous graphs with IP, Port, and Connection nodes. The approach demonstrates strong performance in identifying connection endpoints while maintaining competitive feature reconstruction compared to traditional forecasting baselines.
Network traffic prediction at the flow level is challenging because it requires simultaneously modeling both the graph structure of network connections (which IPs and ports connect to each other) and the dynamic features of those connections. Traditional forecasting approaches may not effectively capture the relational structure inherent in network traffic data, where entities like IPs and ports form a natural graph topology.
The authors use sliding-window techniques to segment network traffic into equal-sized heterogeneous bidirectional graphs containing three types of nodes: IP nodes, Port nodes, and Connection nodes. A Graph Neural Network is then applied to model both the evolution of the graph structure over time and the dynamic features associated with connections. This graph-based representation allows the model to leverage relational information between network entities.
What the paper shows.
The proposed GNN model shows superior results in identifying the Port and IP nodes to which connections attach, outperforming baseline methods on this structural prediction task. Feature reconstruction performance remains competitive with strong forecasting baselines, demonstrating that the model balances both structural and feature-level prediction objectives effectively.
The paper is presented as a proof-of-concept, suggesting limited scope in evaluation. The work lacks detailed comparison with other graph-based or deep learning baselines beyond generic forecasting methods. Scalability to large-scale networks, computational complexity analysis, and evaluation on diverse network datasets are not discussed. The paper does not provide extensive ablation studies or analysis of which components contribute most to performance.
✨ Generated by Claude · Apr 25, 2026 · Read the PDF for authoritative content.