Scholarlink

<p>Click-through rate (CTR) prediction is a critical component of recommender systems, helping infer the likelihood of a user’s engagement (i.e., clicks) with a particular item. Previous studies have typically focused on leveraging either dynamic historical user behaviors or heterogeneous information for feature augmentation. However, relying solely on one aspect is insufficient to capture the intricate user-item dependencies. In this paper, we propose dynamic heterogeneous graph convolutional networks (DH-GCN) for CTR prediction, combining dynamic user-item interactions and heterogeneous information. Specifically, we construct three graphs: item knowledge graph, user-user graph, and user-item graph, and design a novel graph-to-graph learning method to realize the sharing of neighbors and relationships in the GCN framework. Moreover, we leverage multi-granularity time-sliced user-item graphs to capture evolving user preference trajectories. Experiments on three public datasets show that DH-GCN makes significant improvements over state-of-the-art baselines and achieves 0.01-level improvements in AUC, accuracy, and F1.</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

Dynamic heterogeneous graph convolutional networks for click-through rate prediction in recommender systems

  • 期刊论文

  • 发表日期：

    2025年5月18日

查看全文
  • Ying Jin,
  • Yanwu Yang,
  • Baojun Ma

摘要

Click-through rate (CTR) prediction is a critical component of recommender systems, helping infer the likelihood of a user’s engagement (i.e., clicks) with a particular item. Previous studies have typically focused on leveraging either dynamic historical user behaviors or heterogeneous information for feature augmentation. However, relying solely on one aspect is insufficient to capture the intricate user-item dependencies. In this paper, we propose dynamic heterogeneous graph convolutional networks (DH-GCN) for CTR prediction, combining dynamic user-item interactions and heterogeneous information. Specifically, we construct three graphs: item knowledge graph, user-user graph, and user-item graph, and design a novel graph-to-graph learning method to realize the sharing of neighbors and relationships in the GCN framework. Moreover, we leverage multi-granularity time-sliced user-item graphs to capture evolving user preference trajectories. Experiments on three public datasets show that DH-GCN makes significant improvements over state-of-the-art baselines and achieves 0.01-level improvements in AUC, accuracy, and F1.