Scholarlink

<p>Session-based recommendation(SBR) utilizes anonymous sessions to predict the next interaction without considering user profiles. The existing works generally tend to improve SBR performance by learning complex item transitions and enriching session representations. However, the item features obtained by mining the higher-order information do not integrate the contribution of lower-order information, which can easily lead to the over-smoothing problem. Additionally, as a component of session representations, the local interest is typically represented by the last-clicked item of a session, which may deviate from the user’s main intent when the last-clicked one is noise. To solve the abovementioned problems, we propose a novel method named Adaptive Discriminant Feature Learning (ADFL) based on graph neural networks for SBR. The proposed method involves two newly designed components: a paralleled gating module (PGM) and a feature-enhancing unit (FEU). In particular, PGM integrates all orders of item information through the nested application of gating mechanisms, which can overcome the over-smoothing problem and enhance item transitions. Meanwhile, the user’s main intent obtained based on the entire session is introduced to learn the local interest. FEU utilizes the user’s main intent to learn more discriminative and reasonable local interest, easing the effect of the last-clicked item being noise in some sessions. Extensive experiments on three real-world datasets show that the proposed method outperforms the comparative methods. Codes and data are available at <a href="https://github.com/St-ding/ADFL">https://github.com/St-ding/ADFL</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

Adaptive discriminant feature learning for GNN-based session recommendation

  • 期刊论文

  • 发表日期：

    2025年1月4日

查看全文
  • Jiawei Ding,
  • Zhiyi Tan,
  • Guanming Lu,
  • Jinsheng Wei

摘要

Session-based recommendation(SBR) utilizes anonymous sessions to predict the next interaction without considering user profiles. The existing works generally tend to improve SBR performance by learning complex item transitions and enriching session representations. However, the item features obtained by mining the higher-order information do not integrate the contribution of lower-order information, which can easily lead to the over-smoothing problem. Additionally, as a component of session representations, the local interest is typically represented by the last-clicked item of a session, which may deviate from the user’s main intent when the last-clicked one is noise. To solve the abovementioned problems, we propose a novel method named Adaptive Discriminant Feature Learning (ADFL) based on graph neural networks for SBR. The proposed method involves two newly designed components: a paralleled gating module (PGM) and a feature-enhancing unit (FEU). In particular, PGM integrates all orders of item information through the nested application of gating mechanisms, which can overcome the over-smoothing problem and enhance item transitions. Meanwhile, the user’s main intent obtained based on the entire session is introduced to learn the local interest. FEU utilizes the user’s main intent to learn more discriminative and reasonable local interest, easing the effect of the last-clicked item being noise in some sessions. Extensive experiments on three real-world datasets show that the proposed method outperforms the comparative methods. Codes and data are available at https://github.com/St-ding/ADFL.