Scholarlink

<p>Underlying intentions typically drive users’ interactions with items in sequential recommendation. Existing methods for intent-aware sequential recommendation still suffer from two limitations: Existing intent-aware sequential recommendation methods often assume independence among multiple user intents, neglecting implicit interdependencies. Additionally, data sparsity presents challenges for intent modeling and robust user behavior representation, ultimately weakening the integrity of intent representations. To this end, a new sequential recommendation model called <b>F</b>usion of <b>D</b>iffusion Models and <b>I</b>ntent Learning in <b>S</b>equential <b>R</b>ecommendation (FDISR) is proposed. FDISR integrates a novel diffusion-based contrastive Module to generate high-quality interaction sequences for enhanced representation learning. A contrastive multi-head attention Module is introduced to improve the discriminability of intent representations. In contrast, a dynamic intent-aware attention aggregation mechanism captures implicit dependencies among multiple user intents that are often overlooked in prior work. In addition, a mixture-of-experts module is incorporated to enable fine-grained modeling of diverse user interests, allowing the model to capture personalized user interests for final recommendation prediction adaptively. Experiments on three real-world datasets demonstrate that FDISR yields substantial performance improvements, with maximum gains of 25.10% and 38.03% in HR@5 and NDCG@5, respectively.</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

Fusion of diffusion models and intent learning in sequential recommendation

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Jian Feng,
  • Xinyue Jin

摘要

Underlying intentions typically drive users’ interactions with items in sequential recommendation. Existing methods for intent-aware sequential recommendation still suffer from two limitations: Existing intent-aware sequential recommendation methods often assume independence among multiple user intents, neglecting implicit interdependencies. Additionally, data sparsity presents challenges for intent modeling and robust user behavior representation, ultimately weakening the integrity of intent representations. To this end, a new sequential recommendation model called Fusion of Diffusion Models and Intent Learning in Sequential Recommendation (FDISR) is proposed. FDISR integrates a novel diffusion-based contrastive Module to generate high-quality interaction sequences for enhanced representation learning. A contrastive multi-head attention Module is introduced to improve the discriminability of intent representations. In contrast, a dynamic intent-aware attention aggregation mechanism captures implicit dependencies among multiple user intents that are often overlooked in prior work. In addition, a mixture-of-experts module is incorporated to enable fine-grained modeling of diverse user interests, allowing the model to capture personalized user interests for final recommendation prediction adaptively. Experiments on three real-world datasets demonstrate that FDISR yields substantial performance improvements, with maximum gains of 25.10% and 38.03% in HR@5 and NDCG@5, respectively.