Scholarlink

<p>Scientific article recommendation must combine collaborative interaction patterns with the semantic information contained in titles and abstracts. Graph-based collaborative filtering models provide strong warm-start ranking but cannot naturally represent unseen articles, while content-aware methods often underuse graph structure. This paper proposes <b>D</b>isentangled <b>G</b>raph–<b>T</b>ext <b>C</b>ollaborative <b>F</b>iltering (DGT-CF), a framework that separates collaborative and textual representations during learning and connects them only through explicit interfaces, including score-level fusion, a content bridge, orthogonality regularization, and prototype-based cold-start transfer. The collaborative branch is based on LightGCN, while the content branch uses a lightweight dual-stream encoder for article text. A prototype-based cold-start module maps textual features to a mixture of learned collaborative prototypes, enabling zero-shot ranking of new articles. Experiments on CiteULike-A and CiteULike-T, with MIND-small as an additional out-of-domain warm-start validation benchmark, show that DGT-CF achieves consistently strong warm-start performance and improves cold-start ranking over representative collaborative, content-aware, graph contrastive, and language-model baselines.</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

Disentangled graph–text collaborative filtering with prototypical cold-start for scientific article recommendation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月3日

查看全文
  • Jinyang Sun

摘要

Scientific article recommendation must combine collaborative interaction patterns with the semantic information contained in titles and abstracts. Graph-based collaborative filtering models provide strong warm-start ranking but cannot naturally represent unseen articles, while content-aware methods often underuse graph structure. This paper proposes Disentangled Graph–Text Collaborative Filtering (DGT-CF), a framework that separates collaborative and textual representations during learning and connects them only through explicit interfaces, including score-level fusion, a content bridge, orthogonality regularization, and prototype-based cold-start transfer. The collaborative branch is based on LightGCN, while the content branch uses a lightweight dual-stream encoder for article text. A prototype-based cold-start module maps textual features to a mixture of learned collaborative prototypes, enabling zero-shot ranking of new articles. Experiments on CiteULike-A and CiteULike-T, with MIND-small as an additional out-of-domain warm-start validation benchmark, show that DGT-CF achieves consistently strong warm-start performance and improves cold-start ranking over representative collaborative, content-aware, graph contrastive, and language-model baselines.