<p>Graph meta-learning models can fast adapt to new tasks with extremely limited labeled data by learning transferable meta knowledge and inductive bias on graph. Existing methods construct meta-training tasks with abundant labeled nodes from base classes, which limit the application scenarios of graph meta-learning. Therefore, we propose an unsupervised graph meta-learning framework via local subgraph augmentation (<b>UMLGA</b>). Specifically, we firstly propose a graph clustering-based sampling method to sample anchor nodes from different natural classes and extract corresponding local subgraphs. Then, supposing that the generated augmentation samples share the same labels, we design structure-wise and feature-wise graph augmentation strategies to generate diverse augmentation subgraphs while keeping the semantics unchanged. Finally, we perform meta-training on the unsupervised constructed tasks with weighted meta-loss, which can extract cross-tasks knowledge for fast adaption to novel classes. To evaluate the effectiveness of <b>UMLGA</b>, series of experiments are conducted on four real-world graph datasets. Experiment results show that, even without relying on extensive labeled data, <b>UMLGA</b> can achieve comparable and even better few-shot node classification performance comparing with the supervised graph meta-learning backbone models. With GPN as the backbone model, the improvements of <b>UMLGA</b> are respectively 3.0<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\sim \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>∼</mo> </math></EquationSource> </InlineEquation>9.3%, 4.4<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\sim \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>∼</mo> </math></EquationSource> </InlineEquation>11.6%, -1.2<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\sim \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>∼</mo> </math></EquationSource> </InlineEquation>9.3%, and 1.8<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\sim \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>∼</mo> </math></EquationSource> </InlineEquation>15.1% on Amazon-Clothing, Amazon-Electronics, DBLP, and ogbn-products datasets.</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

UMLGA: unsupervised graph meta-learning via local subgraph augmentation

  • Ningbo Huang,
  • Gang Zhou,
  • Meng Zhang,
  • Yi Xia,
  • Shunhang Li

摘要

Graph meta-learning models can fast adapt to new tasks with extremely limited labeled data by learning transferable meta knowledge and inductive bias on graph. Existing methods construct meta-training tasks with abundant labeled nodes from base classes, which limit the application scenarios of graph meta-learning. Therefore, we propose an unsupervised graph meta-learning framework via local subgraph augmentation (UMLGA). Specifically, we firstly propose a graph clustering-based sampling method to sample anchor nodes from different natural classes and extract corresponding local subgraphs. Then, supposing that the generated augmentation samples share the same labels, we design structure-wise and feature-wise graph augmentation strategies to generate diverse augmentation subgraphs while keeping the semantics unchanged. Finally, we perform meta-training on the unsupervised constructed tasks with weighted meta-loss, which can extract cross-tasks knowledge for fast adaption to novel classes. To evaluate the effectiveness of UMLGA, series of experiments are conducted on four real-world graph datasets. Experiment results show that, even without relying on extensive labeled data, UMLGA can achieve comparable and even better few-shot node classification performance comparing with the supervised graph meta-learning backbone models. With GPN as the backbone model, the improvements of UMLGA are respectively 3.0 \(\sim \) 9.3%, 4.4 \(\sim \) 11.6%, -1.2 \(\sim \) 9.3%, and 1.8 \(\sim \) 15.1% on Amazon-Clothing, Amazon-Electronics, DBLP, and ogbn-products datasets.