Scholarlink

<p>Capturing latent intents is essential for enhancing both the accuracy and interpretability of recommender systems. However, existing methods often rely on deep graph propagation and alternating optimization, which fail to explicitly disentangle intents and suffer from limited scalability. To address these challenges, we propose <i>DMI</i>, a unified and end-to-end framework for disentangled multi-intent representation learning. <i>DMI</i> projects user and item behaviors into a shared latent space, where a set of learnable intent prototypes act as semantic anchors to decompose behaviors into coherent intent subspaces. Additionally, <i>DMI</i> incorporates an intent alignment mechanism that captures collaborative intent signals between user and item entities, facilitating semantic alignment across heterogeneous representations. During training, <i>DMI</i> leverages observed user-item interactions as supervisory signals to explicitly guide the alignment between user and item intents, ensuring that the learned representations faithfully capture user preferences and item characteristics. Unlike prior methods that rely on full-graph computation, <i>DMI</i> adopts an efficient node-to-intent message passing scheme with mini-batch training, ensuring scalability to large-scale recommendation scenarios. Extensive experiments demonstrate that <i>DMI</i> not only outperforms state-of-the-art baselines in recommendation accuracy, but also achieves superior intent disentanglement, providing more interpretable and explainable recommendation results.</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

Learning disentangled multi-intent representations for scalable recommendation

  • 期刊论文

  • 发表日期：

    2026年4月29日

查看全文
  • Zhandong Mei,
  • Yongyi Lin,
  • Shanfan Zhang

摘要

Capturing latent intents is essential for enhancing both the accuracy and interpretability of recommender systems. However, existing methods often rely on deep graph propagation and alternating optimization, which fail to explicitly disentangle intents and suffer from limited scalability. To address these challenges, we propose DMI, a unified and end-to-end framework for disentangled multi-intent representation learning. DMI projects user and item behaviors into a shared latent space, where a set of learnable intent prototypes act as semantic anchors to decompose behaviors into coherent intent subspaces. Additionally, DMI incorporates an intent alignment mechanism that captures collaborative intent signals between user and item entities, facilitating semantic alignment across heterogeneous representations. During training, DMI leverages observed user-item interactions as supervisory signals to explicitly guide the alignment between user and item intents, ensuring that the learned representations faithfully capture user preferences and item characteristics. Unlike prior methods that rely on full-graph computation, DMI adopts an efficient node-to-intent message passing scheme with mini-batch training, ensuring scalability to large-scale recommendation scenarios. Extensive experiments demonstrate that DMI not only outperforms state-of-the-art baselines in recommendation accuracy, but also achieves superior intent disentanglement, providing more interpretable and explainable recommendation results.