Scholarlink

<p>The remarkable success of pre-trained large language models (LLMs) in natural language processing has developed a new paradigm of combining LLMs with graph neural networks (GNNs) on modeling textual-attributed graphs. However, manually designing the optimal model architectures to adapt the deep semantics of LLM on different graphs is inefficient and highly demands expert knowledge. Though graph neural architecture search (GNAS) provides a feasible solution to automatically design optimal GNN architectures for different graphs, previous research was mainly established on the shallow embedding methods, which ignored the difference between the deep semantic space and shallow embeddings. Focus on these issues, we propose DSA-GNAS, an automated TAG learning framework based on graph neural architecture search with deep semantic adaptation. To better leverage the deep semantics, we propose a novel structure-semantic fusion (S2F) search space. The model architectures are sampled from the S2F space and form into a dual-path adapter to fine-tune semantic embedding generated by LLM, which can sufficiently make adaptation on the semantic space to the graph downstream task for different TAGs. The model architectures are automatically optimized through a genetic search strategy, which is global and not restricted by gradient, offering promising efficiency in searching for the optimal model architecture. Experimental results show that DSA-GNAS can significantly improve performance on the graph task over other baselines, demonstrating that DSA-GNAS can effectively work in designing optimal model architectures for adapting the deep semantics to graph-related tasks.</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

DSA-GNAS: graph neural architecture search with deep semantic adaptation of large language models

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Siyang Xiao,
  • Jiamin Chen,
  • Zhenpeng Wu,
  • Shuqing Wu,
  • Jianliang Gao

摘要

The remarkable success of pre-trained large language models (LLMs) in natural language processing has developed a new paradigm of combining LLMs with graph neural networks (GNNs) on modeling textual-attributed graphs. However, manually designing the optimal model architectures to adapt the deep semantics of LLM on different graphs is inefficient and highly demands expert knowledge. Though graph neural architecture search (GNAS) provides a feasible solution to automatically design optimal GNN architectures for different graphs, previous research was mainly established on the shallow embedding methods, which ignored the difference between the deep semantic space and shallow embeddings. Focus on these issues, we propose DSA-GNAS, an automated TAG learning framework based on graph neural architecture search with deep semantic adaptation. To better leverage the deep semantics, we propose a novel structure-semantic fusion (S2F) search space. The model architectures are sampled from the S2F space and form into a dual-path adapter to fine-tune semantic embedding generated by LLM, which can sufficiently make adaptation on the semantic space to the graph downstream task for different TAGs. The model architectures are automatically optimized through a genetic search strategy, which is global and not restricted by gradient, offering promising efficiency in searching for the optimal model architecture. Experimental results show that DSA-GNAS can significantly improve performance on the graph task over other baselines, demonstrating that DSA-GNAS can effectively work in designing optimal model architectures for adapting the deep semantics to graph-related tasks.