Scholarlink

<p>Link prediction is an important task in network analysis that aims to estimate missing or future connections between nodes in complex graphs. Recent Graph Neural Network (GNN)-based methods have shown promising performance; however, many existing approaches mainly focus on local neighborhood aggregation and often fail to capture higher-order structural information and community relationships effectively. To overcome these limitations, this paper proposes a Hierarchical Supergraph Neural Network framework for Link Prediction. The proposed framework performs hierarchical representation learning at two levels: node-level learning on the original graph and community-level learning on a community-aware supergraph. The supergraph is constructed by representing communities as supernodes and inter-community relationships as superedges, enabling the model to capture both local and global structural dependencies. The framework integrates Graph Attention Network (GAT)-based node embeddings, community-level representations, node influence information, global graph statistics, and structural heuristic features within a unified architecture. An attention-based fusion mechanism is further employed to combine multi-level graph representations for improved link prediction. Experimental results on several real-world benchmark datasets demonstrate that the proposed framework achieves effective and robust performance compared with existing link prediction methods in terms of AUROC and AUPR.</p>

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

Hierarchical supergraph neural networks for link prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月31日

查看全文
  • Madhusudhana Rao Baswani,
  • Khyathisree Yarra,
  • Y. V. Nandini,
  • Prasanthi Boyapati,
  • T. Jaya Lakshmi

摘要

Link prediction is an important task in network analysis that aims to estimate missing or future connections between nodes in complex graphs. Recent Graph Neural Network (GNN)-based methods have shown promising performance; however, many existing approaches mainly focus on local neighborhood aggregation and often fail to capture higher-order structural information and community relationships effectively. To overcome these limitations, this paper proposes a Hierarchical Supergraph Neural Network framework for Link Prediction. The proposed framework performs hierarchical representation learning at two levels: node-level learning on the original graph and community-level learning on a community-aware supergraph. The supergraph is constructed by representing communities as supernodes and inter-community relationships as superedges, enabling the model to capture both local and global structural dependencies. The framework integrates Graph Attention Network (GAT)-based node embeddings, community-level representations, node influence information, global graph statistics, and structural heuristic features within a unified architecture. An attention-based fusion mechanism is further employed to combine multi-level graph representations for improved link prediction. Experimental results on several real-world benchmark datasets demonstrate that the proposed framework achieves effective and robust performance compared with existing link prediction methods in terms of AUROC and AUPR.