Scholarlink

<p>In recent years, GCN models with heterogeneous graphs as input have been widely used in the field of interpretable recommendation joint tasks. The extraction of explanation information from reviews effectively improves the interpretability and recommendation performance of the recommendation task. However, existing GCN models based on user–item–explanation interaction graphs still have two limitations: (1) Most models ignore the over-smoothing problem in GCNs, which hinders the models from obtaining high-order information from larger neighborhoods to enhance performance. (2) Most models only use the explanation information for the initial embedding construction for item recommendation tasks, which limits the usefulness of explanation information. To address the aforementioned limitations, we propose a new review-aware graph convolutional model named DRE-GCN, which incorporates dual flexible residuals and explanation guidance, for explainable recommendation. Specifically, we design a flexible residual structure controlled by parameters and employ it to construct two independent residual graph convolutional networks for the item recommendation task and explanation ranking task, considering the varying sensitivities of different learning subtasks in joint task model to the over-smoothing problem during graph convolution. Additionally, we propose an explanation-guided aggregation strategy that enables nodes to propagate forward in a biased manner with different importance based on the explanation information. We conducted extensive experiments on four public datasets, and the experimental results demonstrate that DRE-GCN outperforms several state-of-the-art GCN-based models in terms of explanation ranking and item recommendation tasks. The model code is posted to: <a href="https://github.com/ThreeCat-han/DRE-GCN.git.">https://github.com/ThreeCat-han/DRE-GCN.git.</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

DRE-GCN: dual flexible residuals and explanation-guided GCN for explainable recommendation

  • 期刊论文

  • 发表日期：

    2025年12月2日

查看全文
  • Fuqiang Xie,
  • Min Wang,
  • Hongbin Chen,
  • Qi Zhong,
  • Baolei Li

摘要

In recent years, GCN models with heterogeneous graphs as input have been widely used in the field of interpretable recommendation joint tasks. The extraction of explanation information from reviews effectively improves the interpretability and recommendation performance of the recommendation task. However, existing GCN models based on user–item–explanation interaction graphs still have two limitations: (1) Most models ignore the over-smoothing problem in GCNs, which hinders the models from obtaining high-order information from larger neighborhoods to enhance performance. (2) Most models only use the explanation information for the initial embedding construction for item recommendation tasks, which limits the usefulness of explanation information. To address the aforementioned limitations, we propose a new review-aware graph convolutional model named DRE-GCN, which incorporates dual flexible residuals and explanation guidance, for explainable recommendation. Specifically, we design a flexible residual structure controlled by parameters and employ it to construct two independent residual graph convolutional networks for the item recommendation task and explanation ranking task, considering the varying sensitivities of different learning subtasks in joint task model to the over-smoothing problem during graph convolution. Additionally, we propose an explanation-guided aggregation strategy that enables nodes to propagate forward in a biased manner with different importance based on the explanation information. We conducted extensive experiments on four public datasets, and the experimental results demonstrate that DRE-GCN outperforms several state-of-the-art GCN-based models in terms of explanation ranking and item recommendation tasks. The model code is posted to: https://github.com/ThreeCat-han/DRE-GCN.git.