Scholarlink

<p>Real-world e-commerce and media platforms are characterized by sparse feedback, incomplete attributes, and heterogeneous side information, which substantially limit the expressive power of conventional graph-based recommenders. In this work, we propose RichGNN, an attribute-enriched graph neural network (GNN) designed to jointly address interaction sparsity and attribute incompleteness through a unified representation learning framework. By combining structural signals from the interaction graph with enriched semantic information, the proposed model learns a more coherent and expressive latent space that better captures user preferences and item characteristics. Extensive experiments conducted on two widely used benchmarks, MovieLens&#xa0;100K and MovieLens&#xa0;1&#xa0;M, demonstrate the effectiveness of the proposed framework compared to strong baselines achieving consistent and significant improvements, including a relative gain of +5.53% in Recall@20 and +4.64% in NDCG@20 on MovieLens&#xa0;1&#xa0;M, as well as +3.99% in Recall@20 and +2.88% in NDCG@20 on MovieLens&#xa0;100K. Further ablation studies confirm the critical role of attribute enrichment, generative attribute completion, and attention-based fusion in driving these improvements. The results highlight RichGNN as a robust and effective solution for attribute-aware recommendation under sparse and incomplete data conditions.</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

RichGNN: Attribute-enriched graph neural network for optimized e-commerce recommendations

  • 期刊论文

  • 发表日期：

    2026年3月27日

查看全文
  • Abderaouf Bahi,
  • Amel Ourici,
  • Mohamed Amine Ferrag

摘要

Real-world e-commerce and media platforms are characterized by sparse feedback, incomplete attributes, and heterogeneous side information, which substantially limit the expressive power of conventional graph-based recommenders. In this work, we propose RichGNN, an attribute-enriched graph neural network (GNN) designed to jointly address interaction sparsity and attribute incompleteness through a unified representation learning framework. By combining structural signals from the interaction graph with enriched semantic information, the proposed model learns a more coherent and expressive latent space that better captures user preferences and item characteristics. Extensive experiments conducted on two widely used benchmarks, MovieLens 100K and MovieLens 1 M, demonstrate the effectiveness of the proposed framework compared to strong baselines achieving consistent and significant improvements, including a relative gain of +5.53% in Recall@20 and +4.64% in NDCG@20 on MovieLens 1 M, as well as +3.99% in Recall@20 and +2.88% in NDCG@20 on MovieLens 100K. Further ablation studies confirm the critical role of attribute enrichment, generative attribute completion, and attention-based fusion in driving these improvements. The results highlight RichGNN as a robust and effective solution for attribute-aware recommendation under sparse and incomplete data conditions.