Scholarlink

<p>Multimodal recommendation systems improve recommendation performance by leveraging information from different modalities and complex relational networks. However, existing methods still face two main challenges: first, randomly initialized ID embeddings lack semantic and structural priors, making them vulnerable to optimization bias during training, which results in unstable representations and weakens their effectiveness as anchors for multimodal fusion. Second, GCNs typically adopt uniform aggregation for higher-order interactions, which oversmooths node representations and causes semantic degradation. To address these issues, the article proposes a novel multimodal recommendation method based on Structural Activation and Interest-Aware (SAIA). By introducing a perturbation mechanism to enhance the differentiation of ID embeddings, and incorporating gated fusion of modality features with dynamic graph convolution strategies, SAIA effectively solves the issues of homogeneous representations and the identification of node importance. Additionally, we utilize a semantic disentangling mechanism to achieve personalized modality fusion, and employ cross-space contrastive learning to further enhance the semantic compatibility and structural consistency of the fused modality representations. Experimental results show that SAIA outperforms existing methods on four benchmark datasets. Our code is made publicly available on <a href="https://github.com/Blackycat111/SAIA">https://github.com/Blackycat111/SAIA</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

Structure-activated and interest-aware multimodal recommendation method

  • 期刊论文

  • 发表日期：

    2025年12月29日

查看全文
  • HaoYu Wang,
  • HongBin Xia,
  • XiaoFeng Wang

摘要

Multimodal recommendation systems improve recommendation performance by leveraging information from different modalities and complex relational networks. However, existing methods still face two main challenges: first, randomly initialized ID embeddings lack semantic and structural priors, making them vulnerable to optimization bias during training, which results in unstable representations and weakens their effectiveness as anchors for multimodal fusion. Second, GCNs typically adopt uniform aggregation for higher-order interactions, which oversmooths node representations and causes semantic degradation. To address these issues, the article proposes a novel multimodal recommendation method based on Structural Activation and Interest-Aware (SAIA). By introducing a perturbation mechanism to enhance the differentiation of ID embeddings, and incorporating gated fusion of modality features with dynamic graph convolution strategies, SAIA effectively solves the issues of homogeneous representations and the identification of node importance. Additionally, we utilize a semantic disentangling mechanism to achieve personalized modality fusion, and employ cross-space contrastive learning to further enhance the semantic compatibility and structural consistency of the fused modality representations. Experimental results show that SAIA outperforms existing methods on four benchmark datasets. Our code is made publicly available on https://github.com/Blackycat111/SAIA.