Scholarlink

<p>High-value patents are the primary vehicle for driving technological innovation and are a crucial link in the transfer and transformation of patent technology. Identifying high-value patents is critical to overcoming the challenges of patent commercialization. However, existing high-value patent identification methods rarely model complex heterogeneous data, multiple relationships, and the adaptive fusion of heterogeneous relationships. In addition, they focus on single-relationship patent networks and ignore collaborative optimization between different multi-relationship views, resulting in a biased understanding of patent network semantics. To address the above issues, we propose a Multi-Relation Graph Contrastive Enhancement learning framework (MRGCE) for high-value patent identification to model complex patent heterogeneous relations and achieve multiple relations collaborative optimization. First, we construct a complex patent network of patent applicants and inventors based on different meta-path semantics and further combine patent subject technical text to conduct multi-relational network modeling for heterogeneous data. Second, we employ a graph contrast enhancement model to facilitate mutual supervision across different relationship perspectives, collaboratively optimizing the feature propagation and aggregation processes under diverse relationship perspectives, and mining potential commonalities and differences among heterogeneous relationships. Finally, we design a relation-aware multi-head attention mechanism to distinguish the contributions of different relational patent views to high-value patent identification tasks, to adaptively fuse representations of different patent relational patterns, and to conduct an empirical study on a humanoid robot patent dataset. The comparative experimental analysis demonstrates that MRGCE improves the accuracy (ACC) by 3.73%, the recall rate (Recall) by 1.21%, the normalized discounted cumulative gain (NDCG) by 2.16%, and the area under the curve (AUC) also exhibits outstanding performance.</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

MRGCE: a multi-relational graph contrastive enhancement learning framework for high-value patent identification

  • 期刊论文

  • 发表日期：

    2025年8月27日

查看全文
  • Fuxin Wang,
  • Xingshen Liu,
  • Wang Li,
  • Wenfang Tian,
  • Zhixuan Jia,
  • Congjing Ran

摘要

High-value patents are the primary vehicle for driving technological innovation and are a crucial link in the transfer and transformation of patent technology. Identifying high-value patents is critical to overcoming the challenges of patent commercialization. However, existing high-value patent identification methods rarely model complex heterogeneous data, multiple relationships, and the adaptive fusion of heterogeneous relationships. In addition, they focus on single-relationship patent networks and ignore collaborative optimization between different multi-relationship views, resulting in a biased understanding of patent network semantics. To address the above issues, we propose a Multi-Relation Graph Contrastive Enhancement learning framework (MRGCE) for high-value patent identification to model complex patent heterogeneous relations and achieve multiple relations collaborative optimization. First, we construct a complex patent network of patent applicants and inventors based on different meta-path semantics and further combine patent subject technical text to conduct multi-relational network modeling for heterogeneous data. Second, we employ a graph contrast enhancement model to facilitate mutual supervision across different relationship perspectives, collaboratively optimizing the feature propagation and aggregation processes under diverse relationship perspectives, and mining potential commonalities and differences among heterogeneous relationships. Finally, we design a relation-aware multi-head attention mechanism to distinguish the contributions of different relational patent views to high-value patent identification tasks, to adaptively fuse representations of different patent relational patterns, and to conduct an empirical study on a humanoid robot patent dataset. The comparative experimental analysis demonstrates that MRGCE improves the accuracy (ACC) by 3.73%, the recall rate (Recall) by 1.21%, the normalized discounted cumulative gain (NDCG) by 2.16%, and the area under the curve (AUC) also exhibits outstanding performance.