Scholarlink

<p>In recent years, the rapid growth in scholarly publications and the widespread adoption of digital libraries have intensified the demand for effective paper recommendation systems. Traditional approaches typically rely on extensive historical interaction data or rich contextual information from paper abstracts, but often overlook critical temporal dependencies inherent in user preferences. A particular challenge is the recommendation of newly published papers, which, despite their significance in conveying cutting-edge research findings, suffer from sparse historical data. To address these issues, we propose the Meta-path Attention with Semantic Transformer for Academic Recommendation (MAPSTAR) framework, a novel recommendation model that integrates heterogeneous graph attention with transformer-based meta-path attention mechanisms. MAPSTAR simultaneously models both the temporal sequences of user interactions and the complex correlations among papers and their attributes. Specifically, our approach introduces a Transformer Encoder within the meta-path attention layer, allowing each meta-path embedding to capture global dependencies and dynamically adjust its representation based on contextual interactions with other meta-paths.</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

Meta-path attention with semantic transformer for academic recommendation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月19日

查看全文
  • Wei-Cheng Wang,
  • Hai-Yin Huang,
  • Tung-Yang Wu,
  • Ming-Jiu Hwang,
  • Jun-Zhe Wang,
  • Jiun-Long Huang

摘要

In recent years, the rapid growth in scholarly publications and the widespread adoption of digital libraries have intensified the demand for effective paper recommendation systems. Traditional approaches typically rely on extensive historical interaction data or rich contextual information from paper abstracts, but often overlook critical temporal dependencies inherent in user preferences. A particular challenge is the recommendation of newly published papers, which, despite their significance in conveying cutting-edge research findings, suffer from sparse historical data. To address these issues, we propose the Meta-path Attention with Semantic Transformer for Academic Recommendation (MAPSTAR) framework, a novel recommendation model that integrates heterogeneous graph attention with transformer-based meta-path attention mechanisms. MAPSTAR simultaneously models both the temporal sequences of user interactions and the complex correlations among papers and their attributes. Specifically, our approach introduces a Transformer Encoder within the meta-path attention layer, allowing each meta-path embedding to capture global dependencies and dynamically adjust its representation based on contextual interactions with other meta-paths.