Scholarlink

<p>Inventor collaboration networks are both temporally evolving and semantically heterogeneous, involving diverse entity types and complex relational structures. These characteristics present significant challenges for accurately predicting future collaborations. We propose a Dynamic Heterogeneous Graph Neural Network (DHGNN) that jointly captures temporal dynamics and semantic dependencies while enabling interpretable prediction. DHGNN constructs cumulative heterogeneous graph snapshots and integrates relation-aware message passing with meta-path–guided multi-hop aggregation to capture both local and high-order collaboration patterns. A hierarchical attention mechanism combines continuous-time positional encoding with causal masking (preventing information leakage from future timestamps) to dynamically align temporal and semantic signals. For interpretability, DHGNN introduces a two-tier explanation framework: a meta-path attribution module estimates the contribution of each semantic path to the prediction, and a subgraph extraction module visualizes representative local structures aligned with the most influential paths. Experiments on a large-scale inventor–patent dataset demonstrate that DHGNN not only outperforms strong baselines but also generalizes well to cold-start scenarios, such as predicting collaborations between previously unconnected inventors.</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

DHGNN: A dynamic heterogeneous graph neural network for interpretable inventor collaboration prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月13日

查看全文
  • Xiaodong Xie,
  • Jie Wu,
  • Mengjia Xiang,
  • Jianting Tang,
  • Yongxiang Sheng

摘要

Inventor collaboration networks are both temporally evolving and semantically heterogeneous, involving diverse entity types and complex relational structures. These characteristics present significant challenges for accurately predicting future collaborations. We propose a Dynamic Heterogeneous Graph Neural Network (DHGNN) that jointly captures temporal dynamics and semantic dependencies while enabling interpretable prediction. DHGNN constructs cumulative heterogeneous graph snapshots and integrates relation-aware message passing with meta-path–guided multi-hop aggregation to capture both local and high-order collaboration patterns. A hierarchical attention mechanism combines continuous-time positional encoding with causal masking (preventing information leakage from future timestamps) to dynamically align temporal and semantic signals. For interpretability, DHGNN introduces a two-tier explanation framework: a meta-path attribution module estimates the contribution of each semantic path to the prediction, and a subgraph extraction module visualizes representative local structures aligned with the most influential paths. Experiments on a large-scale inventor–patent dataset demonstrate that DHGNN not only outperforms strong baselines but also generalizes well to cold-start scenarios, such as predicting collaborations between previously unconnected inventors.