Scholarlink

Effectively modeling Heterogeneous Information Networks (HINs) is hindered by a dual challenge: the need for both scalable global attention and adaptive, parameter-efficient feature transformation. Existing methods like HGNNs and standard Graph Transformers fail to address both issues simultaneously. To resolve this, we propose MoE-HGT, a Mixture-of-Experts-enhanced Heterogeneous Graph Transformer. Our framework uses sparsely-activated MoE layers to provide specialized, adaptive processing for diverse node types while maintaining parameter efficiency. To achieve scalability, it constructs efficient token sequences from local and metapath-derived contexts, enabling global attention without prohibitive cost. Experiments on four HIN benchmarks show that MoE-HGT consistently outperforms state-of-the-art models.

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

Mixture of Experts Enhanced Heterogeneous Graph Transformer

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Qiheng Mao,
  • Jianling Sun

摘要

Effectively modeling Heterogeneous Information Networks (HINs) is hindered by a dual challenge: the need for both scalable global attention and adaptive, parameter-efficient feature transformation. Existing methods like HGNNs and standard Graph Transformers fail to address both issues simultaneously. To resolve this, we propose MoE-HGT, a Mixture-of-Experts-enhanced Heterogeneous Graph Transformer. Our framework uses sparsely-activated MoE layers to provide specialized, adaptive processing for diverse node types while maintaining parameter efficiency. To achieve scalability, it constructs efficient token sequences from local and metapath-derived contexts, enabling global attention without prohibitive cost. Experiments on four HIN benchmarks show that MoE-HGT consistently outperforms state-of-the-art models.