Scholarlink

<p>News recommendation systems alleviate information overload. Large-scale training and real-time serving involve massive logs and large candidate spaces, creating high-performance computing demands. Existing models face limitations in news semantic representation and user interest modeling. Many methods encode only title or content with a single encoder, failing to capture both local details and global topics, often ignoring the semantic complementarity between click-oriented titles and evidence-rich content. For user modeling, compressing interaction histories into a single vector cannot reflect diverse interests, while candidate-centric attention is sensitive to noisy behaviors. We propose a Dual-Path Bidirectionally-Aware Interest Model (DPBIM). DPBIM enhances news encoding via multi-scale semantic aggregation and title-content interaction. It models users with multiple latent interest subspaces to represent diverse interests and introduces bidirectional historical-candidate attention for matching and noise suppression. Experiments on the MIND benchmark show that DPBIM consistently improves over a range of strong baselines and achieves the best results among the compared methods on AUC, MRR, and nDCG.</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

A news recommendation model with dual-path semantic enhancement and bidirectionally-aware interest modeling

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Shuo Cao,
  • Lili He,
  • Rougang Zhou,
  • Huaxiong Zhang

摘要

News recommendation systems alleviate information overload. Large-scale training and real-time serving involve massive logs and large candidate spaces, creating high-performance computing demands. Existing models face limitations in news semantic representation and user interest modeling. Many methods encode only title or content with a single encoder, failing to capture both local details and global topics, often ignoring the semantic complementarity between click-oriented titles and evidence-rich content. For user modeling, compressing interaction histories into a single vector cannot reflect diverse interests, while candidate-centric attention is sensitive to noisy behaviors. We propose a Dual-Path Bidirectionally-Aware Interest Model (DPBIM). DPBIM enhances news encoding via multi-scale semantic aggregation and title-content interaction. It models users with multiple latent interest subspaces to represent diverse interests and introduces bidirectional historical-candidate attention for matching and noise suppression. Experiments on the MIND benchmark show that DPBIM consistently improves over a range of strong baselines and achieves the best results among the compared methods on AUC, MRR, and nDCG.