Scholarlink

<p>Click-through rate (CTR) is a significant indicator in recommender systems. To make accurate recommendations, effectively mining data that describes user and item features is essential. Conventional machine learning algorithms often show ineffectiveness in this domain. In contrast, recent methods, which are adept at capturing high-order features, have demonstrated superior performance in predicting CTR. These methods generally focus on attention mechanisms and knowledge graph modeling. However, previous models face limitations, such as inadequate modeling of feature dependencies, a lack of flexibility in dynamically adjusting neighbor weights, and overlooking the ability of low-order features to retain historical memory. We propose a new model, the Hierarchical Graph Attention Factorization Machine (Hi-GAFM), to address these issues. This model is trained using a Graph Attention cascade network. The Graph Attention Network can capture feature dependencies flexibly and hierarchically. To capture both low-order and high-order feature information, Hi-GAFM hierarchically inputs weighted features into a feature fusion module, augmenting and fusing these features. Additionally, by utilizing attention scores, Hi-GAFM offers robust model interpretation for click-through rate prediction. We conducted extensive experiments on four real-world datasets. The results demonstrate the proposed model’s effectiveness, efficiency, and superiority over state-of-the-art methods. The resources and implementation of Hi-GAFM are available at: <a href="https://github.com/%2a%2a*">spxwf/Hi-GAFM</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

Hi-GAFM: Hierarchical Interpretable Graph Attention Factorization Machine for CTR Prediction

  • 期刊论文

  • 发表日期：

    2025年10月16日

查看全文
  • Bin Yang,
  • Liusiyuan Sun,
  • Ying Xing,
  • Jinchao Huang,
  • Chen Cheng,
  • Xin Wang

摘要

Click-through rate (CTR) is a significant indicator in recommender systems. To make accurate recommendations, effectively mining data that describes user and item features is essential. Conventional machine learning algorithms often show ineffectiveness in this domain. In contrast, recent methods, which are adept at capturing high-order features, have demonstrated superior performance in predicting CTR. These methods generally focus on attention mechanisms and knowledge graph modeling. However, previous models face limitations, such as inadequate modeling of feature dependencies, a lack of flexibility in dynamically adjusting neighbor weights, and overlooking the ability of low-order features to retain historical memory. We propose a new model, the Hierarchical Graph Attention Factorization Machine (Hi-GAFM), to address these issues. This model is trained using a Graph Attention cascade network. The Graph Attention Network can capture feature dependencies flexibly and hierarchically. To capture both low-order and high-order feature information, Hi-GAFM hierarchically inputs weighted features into a feature fusion module, augmenting and fusing these features. Additionally, by utilizing attention scores, Hi-GAFM offers robust model interpretation for click-through rate prediction. We conducted extensive experiments on four real-world datasets. The results demonstrate the proposed model’s effectiveness, efficiency, and superiority over state-of-the-art methods. The resources and implementation of Hi-GAFM are available at: spxwf/Hi-GAFM.