Scholarlink

Dynamic Link Prediction (DLP) aims to forecast the formation or disappearance of links in evolving networks. Traditional GNN-based models (e.g., GCN-GRU, GAT-GRU) effectively capture localized spatial-temporal dependencies but often fall short in modeling long-range structural and temporal patterns. Recent advancements in Graph Transformers introduce global attention mechanisms that better capture such dependencies. This paper presents a benchmarking study of six representative models—RoleGRU, GCN-GRU, GAT-GRU, Transformer, GraphTransformer-GRU, and a hybrid GCN+Transformer evaluated on two real-world datasets: MOOC and Enron. Results show that GraphTransformer-GRU consistently achieves the highest AUC and Average Precision, outperforming both recurrent and hybrid designs. These findings offer key insights into architectural trade-offs and reinforce the importance of integrating spatial structure and temporal dynamics in DLP, addressing gaps noted in recent studies.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Benchmarking GNN and Graph Transformer Models for Dynamic Link Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nahid Abdolrahmanpour Holagh

摘要

Dynamic Link Prediction (DLP) aims to forecast the formation or disappearance of links in evolving networks. Traditional GNN-based models (e.g., GCN-GRU, GAT-GRU) effectively capture localized spatial-temporal dependencies but often fall short in modeling long-range structural and temporal patterns. Recent advancements in Graph Transformers introduce global attention mechanisms that better capture such dependencies. This paper presents a benchmarking study of six representative models—RoleGRU, GCN-GRU, GAT-GRU, Transformer, GraphTransformer-GRU, and a hybrid GCN+Transformer evaluated on two real-world datasets: MOOC and Enron. Results show that GraphTransformer-GRU consistently achieves the highest AUC and Average Precision, outperforming both recurrent and hybrid designs. These findings offer key insights into architectural trade-offs and reinforce the importance of integrating spatial structure and temporal dynamics in DLP, addressing gaps noted in recent studies.