Scholarlink

<p>Graph transformers (GTs) exhibit significant potential in graph representation learning (GRL), yet their performance is critically dependent on the effective encoding of graph topological structures. Existing methods employing a single positional encoding (PE) offer limited perspectives, while fixed PE combinations lack the necessary flexibility. To address these challenges, this work introduces AutoFusion, an adaptive framework for the fusion and selection of PEs. AutoFusion integrates a diverse set of heterogeneous PEs, derived from spectral methods, diffusion kernels, relative distances, and substructure patterns, to provide rich structural insights. Importantly, it incorporates a dual-level adaptive selection mechanism. During an efficient warmup phase on a data subset early in model training, AutoFusion utilizes learnable gates to concurrently optimize both the importance weights of different PE types and the selection weights for various fusion strategies. Upon completion of this warmup phase, the optimal PE subset and fusion strategy are fixed for subsequent, efficient model training. This automated mechanism transforms the computationally intensive architectural search for an optimal structural information scheme into an efficient optimization task. It thereby obviates the need for laborious manual tuning that would otherwise require high-performance computing environments, while ensuring the model leverages the most pertinent combination of structural information in a data-driven manner. We conduct extensive evaluations across multiple benchmarks, encompassing node classification, graph classification and graph regression tasks. Across a range of tasks including node classification, graph classification, and graph regression, AutoFusion-GT significantly outperforms strong baselines that employ single PEs or fixed fusion strategies. Ablation studies further corroborate the necessity of multi-PE fusion and underscore the effectiveness of AutoFusion’s dual selection mechanism and efficient warmup strategy.</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

Autofusion: an adaptive multi-positional encoding fusion framework for graph transformers

  • 期刊论文

  • 发表日期：

    2025年10月31日

查看全文
  • Tianming Yu,
  • Liyu Guan,
  • Qiliang Zhang,
  • Zongwei Hu

摘要

Graph transformers (GTs) exhibit significant potential in graph representation learning (GRL), yet their performance is critically dependent on the effective encoding of graph topological structures. Existing methods employing a single positional encoding (PE) offer limited perspectives, while fixed PE combinations lack the necessary flexibility. To address these challenges, this work introduces AutoFusion, an adaptive framework for the fusion and selection of PEs. AutoFusion integrates a diverse set of heterogeneous PEs, derived from spectral methods, diffusion kernels, relative distances, and substructure patterns, to provide rich structural insights. Importantly, it incorporates a dual-level adaptive selection mechanism. During an efficient warmup phase on a data subset early in model training, AutoFusion utilizes learnable gates to concurrently optimize both the importance weights of different PE types and the selection weights for various fusion strategies. Upon completion of this warmup phase, the optimal PE subset and fusion strategy are fixed for subsequent, efficient model training. This automated mechanism transforms the computationally intensive architectural search for an optimal structural information scheme into an efficient optimization task. It thereby obviates the need for laborious manual tuning that would otherwise require high-performance computing environments, while ensuring the model leverages the most pertinent combination of structural information in a data-driven manner. We conduct extensive evaluations across multiple benchmarks, encompassing node classification, graph classification and graph regression tasks. Across a range of tasks including node classification, graph classification, and graph regression, AutoFusion-GT significantly outperforms strong baselines that employ single PEs or fixed fusion strategies. Ablation studies further corroborate the necessity of multi-PE fusion and underscore the effectiveness of AutoFusion’s dual selection mechanism and efficient warmup strategy.