Scholarlink

<p>Conversational recommender system (CRS) captures user preferences from historical conversations to provide high-quality recommendations. Existing studies introduce external data to compensate for data limitations, improving the model’s ability to predict user preferences. However, CRS still faces several challenges: i) Existing methods only use text features to capture user preferences, ignoring the role of images in entity representation. The fusion mechanism in CRS fails to effectively enable parameter interaction between multiple sources, leading to lower-quality feature fusion results. ii) In the recommendation module, previous methods assign weights to entities in a sequential manner but overlook the influence of contextual semantic information over time. This makes it difficult to update entity weights in real-time. iii) In the dialogue module, the system-generated responses lack diversity. To address these issues, this paper first introduces entity-related image information to enrich entity features, encoding relevant themes and storylines of items. Second, to reduce differences between multiple sources, we propose a Cross Correlation Interaction Mechanism (CCIM) to enable effective interaction and achieve efficient feature fusion. Finally, we incorporate diffusion models into CRS. In the recommendation module, the diffusion model updates entity weights dynamically through iterative learning, capturing the dynamic evolution of user interests. In the dialogue module, we design an entity diffusion module to improve response diversity. Extensive experiments on two public CRS datasets demonstrate the effectiveness of our model. Our code is released on <a href="https://github.com/Janns0916/MDCR-main">https://github.com/Janns0916/MDCR-main</a><a href="https://github.com/Janns0916/MDCR">https://github.com/Janns0916/MDCR</a>.</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-source information diffusion model for conversation recommender system

  • 期刊论文

  • 发表日期：

    2025年5月17日

查看全文
  • Fengjin Liu,
  • Xianying Huang,
  • Huaiyu Liu

摘要

Conversational recommender system (CRS) captures user preferences from historical conversations to provide high-quality recommendations. Existing studies introduce external data to compensate for data limitations, improving the model’s ability to predict user preferences. However, CRS still faces several challenges: i) Existing methods only use text features to capture user preferences, ignoring the role of images in entity representation. The fusion mechanism in CRS fails to effectively enable parameter interaction between multiple sources, leading to lower-quality feature fusion results. ii) In the recommendation module, previous methods assign weights to entities in a sequential manner but overlook the influence of contextual semantic information over time. This makes it difficult to update entity weights in real-time. iii) In the dialogue module, the system-generated responses lack diversity. To address these issues, this paper first introduces entity-related image information to enrich entity features, encoding relevant themes and storylines of items. Second, to reduce differences between multiple sources, we propose a Cross Correlation Interaction Mechanism (CCIM) to enable effective interaction and achieve efficient feature fusion. Finally, we incorporate diffusion models into CRS. In the recommendation module, the diffusion model updates entity weights dynamically through iterative learning, capturing the dynamic evolution of user interests. In the dialogue module, we design an entity diffusion module to improve response diversity. Extensive experiments on two public CRS datasets demonstrate the effectiveness of our model. Our code is released on https://github.com/Janns0916/MDCR-mainhttps://github.com/Janns0916/MDCR.