Scholarlink

<p>Forecasting the evolution of research topics is essential for identifying emerging trends in rapidly evolving scientific domains. In this study, scholarly themes are modeled as temporal keyword co-occurrence networks to capture their structural and temporal dynamics. A multi-domain evaluation is conducted using three graphs derived from the DBLP Citation Network V14: the Cyber Security Research Graph (CSRG), the Artificial Intelligence Research Graph (AIRG), and the Social Network Research Graph (SNRG), each represented as yearly snapshots (2001–2022). The datasets exhibit high temporal novelty, where most keyword associations emerge dynamically over time, motivating the use of inductive models. A GraphSAGE-based framework is employed for dynamic link prediction, evaluating multiple aggregation functions and neighborhood sampling strategies under a temporal inductive setting. Experimental results (2018–2022) demonstrate strong predictive performance across all datasets. Importantly, the findings reveal that no single configuration is universally optimal; instead, model performance is strongly influenced by graph structural properties such as size, density, and local connectivity. These results highlight the importance of structure-aware model design and demonstrate the effectiveness of inductive graph neural networks for forecasting evolving keyword relationships in dynamic scholarly environments.</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

Evaluating aggregators and sampling strategies in GraphSAGE for dynamic link prediction in keyword co-occurrence network

  • 期刊论文

  • 发表日期：

    2026年4月26日

查看全文
  • Anab Batool Kazmi,
  • Muhammad Arshad Islam

摘要

Forecasting the evolution of research topics is essential for identifying emerging trends in rapidly evolving scientific domains. In this study, scholarly themes are modeled as temporal keyword co-occurrence networks to capture their structural and temporal dynamics. A multi-domain evaluation is conducted using three graphs derived from the DBLP Citation Network V14: the Cyber Security Research Graph (CSRG), the Artificial Intelligence Research Graph (AIRG), and the Social Network Research Graph (SNRG), each represented as yearly snapshots (2001–2022). The datasets exhibit high temporal novelty, where most keyword associations emerge dynamically over time, motivating the use of inductive models. A GraphSAGE-based framework is employed for dynamic link prediction, evaluating multiple aggregation functions and neighborhood sampling strategies under a temporal inductive setting. Experimental results (2018–2022) demonstrate strong predictive performance across all datasets. Importantly, the findings reveal that no single configuration is universally optimal; instead, model performance is strongly influenced by graph structural properties such as size, density, and local connectivity. These results highlight the importance of structure-aware model design and demonstrate the effectiveness of inductive graph neural networks for forecasting evolving keyword relationships in dynamic scholarly environments.