[B-7] 情報ネットワーク

11:00 〜 11:15キャリアエクスプローラー

[B-7-09] Scalability Analysis of GNN-Based Network Resilience Evaluation under Topology-Size Shift

△Tianhao Zhu¹, Hiroyuki Morikawa¹, Yoshiaki Narusue¹ (1. The Univ. of Tokyo)

キーワード：

GNN、Network resilience、Topology-size shift

This paper analyzes topology-size shift in GNN-based resilience evaluation for OSPF networks. We formulate packet-loss prediction under link failures as a graph-level binary classification problem and evaluate models on larger unseen topologies. We compare multiple GNN backbones and positional encodings using F1 score and recall curves over topology size. Experimental results show that baseline models degrade at different rates under topology-size shift, while the proposed MPNN with resilience-aware positional encoding achieves the smallest F1-score drop. These results indicate that performance curves over topology size are useful for evaluating model scalability and that resilience-aware structural information improves robustness across different topology sizes.