Scholarlink

<p>Multimodal recommendation systems leverage rich auxiliary information to enhance performance. However, effectively fusing heterogeneous modalities while suppressing noise remains a significant challenge. To address this, we propose a diffusion-based contrastive learning framework for multimodal recommendation (DiffCLR). To overcome the structural bottlenecks of static graph propagation, DiffCLR introduces a multimodal graph denoising diffusion probabilistic model (MG-DDPM) for effective preference disentanglement. This component generates modality-specific interaction graphs by integrating a dynamic gating mechanism and a preference alignment constraint into the reverse denoising process. This generative paradigm allows the model to isolate intrinsic modality-dependent interests from noisy interactions. Additionally, we design a dual-path propagation architecture and a cross-view contrastive learning strategy to enhance the robustness and semantic alignment of the learned embeddings. Extensive experiments on four public benchmarks demonstrate that DiffCLR consistently outperforms state-of-the-art baselines, achieving an average improvement of 6.07%. The code for this work is available at <a href="https://github.com/sun2ot/DiffCLR">https://github.com/sun2ot/DiffCLR</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

Diffusion-based contrastive learning for multimodal recommendation

  • 期刊论文

  • 发表日期：

    2026年3月26日

查看全文
  • Hairong Wang,
  • Zhihang Yi,
  • Zhaojing Xu,
  • Jianling Yang

摘要

Multimodal recommendation systems leverage rich auxiliary information to enhance performance. However, effectively fusing heterogeneous modalities while suppressing noise remains a significant challenge. To address this, we propose a diffusion-based contrastive learning framework for multimodal recommendation (DiffCLR). To overcome the structural bottlenecks of static graph propagation, DiffCLR introduces a multimodal graph denoising diffusion probabilistic model (MG-DDPM) for effective preference disentanglement. This component generates modality-specific interaction graphs by integrating a dynamic gating mechanism and a preference alignment constraint into the reverse denoising process. This generative paradigm allows the model to isolate intrinsic modality-dependent interests from noisy interactions. Additionally, we design a dual-path propagation architecture and a cross-view contrastive learning strategy to enhance the robustness and semantic alignment of the learned embeddings. Extensive experiments on four public benchmarks demonstrate that DiffCLR consistently outperforms state-of-the-art baselines, achieving an average improvement of 6.07%. The code for this work is available at https://github.com/sun2ot/DiffCLR.