Scholarlink

<p>Accurately forecasting research trends is essential for strategic decision-making in academia, funding allocation, and scientific foresight. Traditional citation-based approaches are limited by the slow accumulation of citations and disciplinary biases, which hinder the timely detection of emerging topics. This study proposes a method to predict academic keyword trends using temporal frequency patterns of author-defined keywords as a more immediate and responsive indicator of topical momentum. A novel dataset, the DBLP-Temporal Keyword Trend, is introduced. It comprises year-wise keyword frequency data extracted from the DBLP Citation Network V14. Two deep learning models are developed, DS-GRU-1Y and DS-GRU-5Y, based on deep-stacked gated recurrent unit architectures, for short-term and long-term trend prediction, respectively. The proposed models outperform the baseline recurrent neural network variants in multiple evaluation metrics. A rule-based trend classification algorithm further categorizes keywords into Always Rising, Always Declining, and Mixed patterns based on their predicted trajectories. This keyword-centric framework facilitates the early identification of emerging research topics and provides a scalable, interpretable tool for research planning, topical monitoring, and bibliometric insight.</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

DBLP-TKT: a time-series dataset and deep learning model for academic keyword trend prediction

  • 期刊论文

  • 发表日期：

    2026年2月10日

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

摘要

Accurately forecasting research trends is essential for strategic decision-making in academia, funding allocation, and scientific foresight. Traditional citation-based approaches are limited by the slow accumulation of citations and disciplinary biases, which hinder the timely detection of emerging topics. This study proposes a method to predict academic keyword trends using temporal frequency patterns of author-defined keywords as a more immediate and responsive indicator of topical momentum. A novel dataset, the DBLP-Temporal Keyword Trend, is introduced. It comprises year-wise keyword frequency data extracted from the DBLP Citation Network V14. Two deep learning models are developed, DS-GRU-1Y and DS-GRU-5Y, based on deep-stacked gated recurrent unit architectures, for short-term and long-term trend prediction, respectively. The proposed models outperform the baseline recurrent neural network variants in multiple evaluation metrics. A rule-based trend classification algorithm further categorizes keywords into Always Rising, Always Declining, and Mixed patterns based on their predicted trajectories. This keyword-centric framework facilitates the early identification of emerging research topics and provides a scalable, interpretable tool for research planning, topical monitoring, and bibliometric insight.