Scholarlink

<p>Recommender systems’ performance hinges on dataset quality, but implicit feedback is pervaded by noise that degrades accuracy. Existing denoising methods either rely on unavailable prior knowledge or oversimplify noise via fixed rules. We propose MADR, an enhanced method fusing multi-head attention and adaptive weighting that refines the reweighting paradigm. It integrates cross-epoch loss patterns and user–item semantics to estimate noise probabilities, dynamically adjusts sample weights, and adopts a popularity-sensitive negative sampling strategy to mitigate bias. Experiments on three datasets show that MADR achieves the best or competitive performance on most key metrics, with an average improvement of 2.49% over state-of-the-art methods. Leveraging HPC capabilities enables its scalable deployment on billion-scale datasets, bridging denoising effectiveness and real-world applicability.</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-head attention and dynamic weight adjustment-based denoising method for recommender systems

  • 期刊论文

  • 发表日期：

    2026年3月27日

查看全文
  • Yuqiang Li,
  • Dengshuai Zhang,
  • Lin Li

摘要

Recommender systems’ performance hinges on dataset quality, but implicit feedback is pervaded by noise that degrades accuracy. Existing denoising methods either rely on unavailable prior knowledge or oversimplify noise via fixed rules. We propose MADR, an enhanced method fusing multi-head attention and adaptive weighting that refines the reweighting paradigm. It integrates cross-epoch loss patterns and user–item semantics to estimate noise probabilities, dynamically adjusts sample weights, and adopts a popularity-sensitive negative sampling strategy to mitigate bias. Experiments on three datasets show that MADR achieves the best or competitive performance on most key metrics, with an average improvement of 2.49% over state-of-the-art methods. Leveraging HPC capabilities enables its scalable deployment on billion-scale datasets, bridging denoising effectiveness and real-world applicability.