Scholarlink

<p>Sequential Recommendation (SR) aims to predict the next item a user will interact with based on their historical behaviors. However, real-world interaction sequences often contain irrelevant items, and existing models generally fail to filter out these interactions, thereby propagating noisy contextual information that degrades recommendation accuracy. To address this issue, we propose MDR, a novel Multi-step Denoising Retrieval framework for sequential recommendation. MDR&#xa0;designs a selective denoising Transformer block that progressively suppresses noisy patterns and captures denoised sequential dependencies. To further enhance representation quality, we design a multi-step retrieval paradigm equipped with an Mixture-of-Experts (MoE)-based fusing module, which adaptively integrates fine-grained temporal representations through a gating mechanism. This design not only strengthens contextual modeling but also improves robustness against noisy or perturbed interactions. Extensive experiments on four real-world datasets demonstrate that MDR&#xa0;consistently outperforms state-of-the-art sequential recommendation methods in terms of both effectiveness and robustness, validating the superiority of our denoising-driven design.</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

Multi-step denoising retrieval with progressive mixture-of-experts for sequential recommendation

  • 期刊论文

  • 发表日期：

    2026年1月19日

查看全文
  • Yuhua Mo,
  • Xinglin Lian,
  • Chao Deng,
  • Jibin Wang,
  • Xipeng Liu,
  • Paiyu Li,
  • Liying Huang,
  • Yuhao Ma,
  • Fan Zhou

摘要

Sequential Recommendation (SR) aims to predict the next item a user will interact with based on their historical behaviors. However, real-world interaction sequences often contain irrelevant items, and existing models generally fail to filter out these interactions, thereby propagating noisy contextual information that degrades recommendation accuracy. To address this issue, we propose MDR, a novel Multi-step Denoising Retrieval framework for sequential recommendation. MDR designs a selective denoising Transformer block that progressively suppresses noisy patterns and captures denoised sequential dependencies. To further enhance representation quality, we design a multi-step retrieval paradigm equipped with an Mixture-of-Experts (MoE)-based fusing module, which adaptively integrates fine-grained temporal representations through a gating mechanism. This design not only strengthens contextual modeling but also improves robustness against noisy or perturbed interactions. Extensive experiments on four real-world datasets demonstrate that MDR consistently outperforms state-of-the-art sequential recommendation methods in terms of both effectiveness and robustness, validating the superiority of our denoising-driven design.