[论文解读] CoCo: A Coupled Contrastive Framework for Unsupervised Domain Adaptive Graph Classification

Nan Yin, Li Shen+6 more|arXiv (Cornell University)|Jun 8, 2023

Advanced Graph Neural Networks被引用 9

一句话总结

CoCo 引入一个双分支框架（GCN 和 hierarchical graph kernel network）并结合耦合和跨域对比学习，以实现无监督域自适应图分类。它在多样的图数据集上展示了更好的迁移效果。

ABSTRACT

Although graph neural networks (GNNs) have achieved impressive achievements in graph classification, they often need abundant task-specific labels, which could be extensively costly to acquire. A credible solution is to explore additional labeled graphs to enhance unsupervised learning on the target domain. However, how to apply GNNs to domain adaptation remains unsolved owing to the insufficient exploration of graph topology and the significant domain discrepancy. In this paper, we propose Coupled Contrastive Graph Representation Learning (CoCo), which extracts the topological information from coupled learning branches and reduces the domain discrepancy with coupled contrastive learning. CoCo contains a graph convolutional network branch and a hierarchical graph kernel network branch, which explore graph topology in implicit and explicit manners. Besides, we incorporate coupled branches into a holistic multi-view contrastive learning framework, which not only incorporates graph representations learned from complementary views for enhanced understanding, but also encourages the similarity between cross-domain example pairs with the same semantics for domain alignment. Extensive experiments on popular datasets show that our CoCo outperforms these competing baselines in different settings generally.

研究动机与目标

• 在目标标签稀缺或缺失的情况下，推动图分类的无监督域自适应。
• 提出一个双分支架构，以捕捉隐式和显式的图拓扑信息。
• 开发一个多视角对比学习框架，使跨分支和跨域的表示对齐。
• 在标准图分类基准数据集上验证该方法并分析消融实验。

提出的方法

• 两个耦合分支：一个用于隐式拓扑学习的 Graph Convolutional Network (GCN) 分支。
• 一个使用可学习图滤波器和图核来显式捕获拓扑的 Hierarchical Graph Kernel Network (HGKN) 分支。
• 跨分支对比学习，用于对齐每个图的 GCN 与 HGKN 表示。
• 通过伪标签进行跨域对比学习，以通过 EM 类似的似然框架减少域差异。
• 一个将跨分支、跨域和监督损失结合在一起的整体目标函数。
• 实现中，在 GCN 分支采用两层 GIN，在 HGKN 中使用基于 WL 核的分支，带可训练滤波器和量化，以实现端到端学习。

Figure 1: An overview of the proposed CoCo. Our CoCo feeds source graphs and target graphs into two branches (i.e., GCN branch and HGKN branch). Cross-branch contrastive learning compares graph representations from two branches to achieve while cross-branch contrastive learning compares graph repres

实验结果

研究问题

• RQ1 如何在标记数据有限的情况下有效提取并利用图的拓扑信息？
• RQ2 将隐式（GCN）和显式（HGKN）拓扑表示耦合，是否能在域迁移下提升图分类性能？
• RQ3 通过伪标签引导的跨域对比学习，是否能在语义上对齐源域和目标域的图以提升域自适应？
• RQ4 提出的基于 EM 的视角对跨域对比目标是否有效并改善域对齐？

主要发现

• CoCo 在多项迁移任务和数据集上优于广泛的基线方法（包括 GCN、WL 子树、各种 DA 方法）。
• 跨分支对比学习通过结合互补的拓扑视图来改善图表示。
• 通过伪标签的跨域对比学习通过拉近同语义的跨域对来降低域差异。
• 消融研究表明移除跨分支或跨域对比成分会降低性能，凸显了它们的重要性。
• 该框架在 Mutagenicity 等基准上实现了显著的平均提升，结果显示出强鲁棒的域自适应能力。

Figure 2: An illustration of the hierarchical graph kernel network branch. HGKN compares each $r$ -hop subgraph with filter graphs to generate node features, followed by an attention-based pooling operation to coarsen the graph.