International Conference on Information and Knowledge Technology

فارسی

Home / شانزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

GNN-based Topology Feature Extraction for Adaptive 6G Network Slicing

Authors :

Amirmasoud Sepehrian¹ Siavash Khorsandi²

1- دانشگاه صنعتی امیرکبیر (پلی‌تکنیک تهران) 2- دانشگاه صنعتی امیرکبیر (پلی‌تکنیک تهران)

Keywords :

6G Networks،Soft Network Slicing،Graph Neural Networks،Topology Feature Extraction،Representation Power،Comparative Evaluation

Abstract :

The evolution to 6G networks introduces unprecedented challenges, including ultra-high data rates, massive connectivity, and stringent QoS demands (e.g., sub-millisecond latency for URLLC) in highly dynamic, heterogeneous environments. Traditional hard slicing methods fall short in adapting to fluctuating traffic and resource availability, leading to inefficiencies in resource utilization, SLA violations, and increased energy consumption. This necessitates advanced adaptive mechanisms like soft network slicing, which require precise topology descriptions to predict performance metrics and enable real-time orchestration. Graph Neural Networks (GNNs) are essential here, as they excel at capturing intricate graph-structured relationships in network topologies—far superior to conventional ML models that ignore relational dependencies—facilitating scalable feature extraction for optimization tasks. This research addresses these needs through two core components: (1) a comprehensive comparison of GNN variants (GraphSAGE, GCN, GAT, TransformerConv) to evaluate their representation power in terms of descriptive accuracy and runtime; and (2) a novel embedding method that integrates current slicing requests and global graph features (e.g., density, centrality) with local attributes. Using the Internet Topology Zoo dataset augmented with 6G slice variants, we assess models on metrics like MSE, R2, SMAPE, runtime efficiency, and generalization.