[论文解读] Hypergraph-enhanced Dual Semi-supervised Graph Classification

Wei Ju, Zhengyang Mao+7 more|arXiv (Cornell University)|May 8, 2024

Advanced Graph Neural Networks被引用 8

一句话总结

HEAL 引入基于超图和线图的双框架，用于半监督图分类，利用可学习的超图结构和关系一致性学习来利用未标记图并捕捉高阶依赖。它在若干基准数据集上实现了最先进的性能。

ABSTRACT

In this paper, we study semi-supervised graph classification, which aims at accurately predicting the categories of graphs in scenarios with limited labeled graphs and abundant unlabeled graphs. Despite the promising capability of graph neural networks (GNNs), they typically require a large number of costly labeled graphs, while a wealth of unlabeled graphs fail to be effectively utilized. Moreover, GNNs are inherently limited to encoding local neighborhood information using message-passing mechanisms, thus lacking the ability to model higher-order dependencies among nodes. To tackle these challenges, we propose a Hypergraph-Enhanced DuAL framework named HEAL for semi-supervised graph classification, which captures graph semantics from the perspective of the hypergraph and the line graph, respectively. Specifically, to better explore the higher-order relationships among nodes, we design a hypergraph structure learning to adaptively learn complex node dependencies beyond pairwise relations. Meanwhile, based on the learned hypergraph, we introduce a line graph to capture the interaction between hyperedges, thereby better mining the underlying semantic structures. Finally, we develop a relational consistency learning to facilitate knowledge transfer between the two branches and provide better mutual guidance. Extensive experiments on real-world graph datasets verify the effectiveness of the proposed method against existing state-of-the-art methods.

研究动机与目标

• 在标记图稀少、未标记图丰富的情况下，激发半监督图分类的动机。
• 通过可学习的超图捕捉高阶节点依赖，超越局部邻域。
• 在学习到的超边上，通过线图建模高阶子结构之间的交互。
• 通过关系一致性学习，在超图分支与线图分支之间实现相互知识传递。
• 在真实世界数据集上展示相对于最前沿基线算法的鲁棒性和有效性。

提出的方法

• 引入可学习的超图结构，以含低秩超边矩阵 Λ = H W (k 条超边) 捕捉更高阶的节点依赖。
• 应用超图卷积以获得高阶节点表示和图级嵌入 s_G。
• 在学习到的超图上构建线图，使用 GNN 编码器获得第二个图级嵌入 t_G。
• 提出关系一致性学习，通过内存库中的锚点图对齐超图视图和线图视图之间的相似性分布。
• 在带标签的图上使用有监督损失进行训练，并对未标记图施加关系一致性损失以实现正则化。

实验结果

研究问题

• RQ1 学习一个超图结构如何改进对图分类中高阶节点交互的建模？
• RQ2 在学习到的超边上的线图是否能捕捉超图视图之外的超边之间的交互与语义？
• RQ3 关系一致性学习是否能有效利用未标记图提升半监督性能？
• RQ4 与最先进的半监督图分类方法相比，HEAL 的整体性能提升有多大？

主要发现

方法	PROTEINS	DD	IMDB-B	IMDB-M	REDDIT-M-5k	COLLAB
WL	63.5 ± 1.6	57.3 ± 1.2	58.1 ± 2.3	33.3 ± 1.4	37.0 ± 0.9	62.9 ± 0.7
Sub2Vec	52.7 ± 4.5	46.4 ± 3.2	44.9 ± 3.5	31.8 ± 2.7	35.1 ± 1.5	60.8 ± 1.4
Graph2Vec	63.1 ± 1.8	53.7 ± 1.6	61.2 ± 2.6	38.1 ± 2.2	38.1 ± 1.4	63.6 ± 0.9
EntMin	62.7 ± 2.7	59.8 ± 1.3	67.1 ± 3.7	37.4 ± 1.2	38.7 ± 2.8	63.8 ± 1.6
Mean-Teacher	64.3 ± 2.1	60.6 ± 1.8	66.4 ± 2.7	38.8 ± 3.6	39.2 ± 2.1	63.6 ± 1.4
VAT	64.1 ± 1.2	59.9 ± 2.6	67.2 ± 2.9	39.6 ± 1.4	38.9 ± 3.2	64.1 ± 1.1
InfoGraph	68.2 ± 0.7	67.5 ± 1.4	71.8 ± 2.3	42.3 ± 1.8	41.5 ± 1.7	65.7 ± 0.4
ASGN	67.7 ± 1.2	68.5 ± 0.6	70.6 ± 1.4	41.2 ± 1.4	42.2 ± 0.8	65.3 ± 0.8
GraphCL	69.4 ± 0.8	68.7 ± 1.2	71.2 ± 2.5	43.7 ± 1.3	42.3 ± 0.9	66.4 ± 0.6
JOAO	68.7 ± 0.9	67.9 ± 1.3	71.0 ± 1.9	42.6 ± 1.5	42.1 ± 1.2	65.8 ± 0.4
DualGraph	70.1 ± 1.2	69.8 ± 0.8	72.1 ± 0.7	44.8 ± 0.4	42.9 ± 1.4	67.2 ± 0.6
KGNN	70.9 ± 0.5	70.5 ± 0.6	72.5 ± 1.6	43.3 ± 0.7	44.8 ± 0.6	67.4 ± 0.5
TGNN	71.0 ± 0.7	70.8 ± 0.9	72.8 ± 1.7	42.9 ± 0.8	43.8 ± 1.0	67.7 ± 0.4
HEAL	73.4 ± 0.8	72.1 ± 0.9	73.5 ± 1.5	44.3 ± 0.6	45.9 ± 1.0	68.3 ± 0.5

• HEAL 在六个基准数据集上的多项指标上超越基线，在若干任务上达到已报道的最佳分数。
• 消融研究表明，将超图视图与线图视图结合并结合关系一致性，可获得最强的性能。
• 关系一致性学习有效利用未标记图，在鲁棒性和准确性方面优于纯监督或单分支变体。
• 超参数：嵌入维度 d=32 和超边数 k=32 在性能与效率之间提供了有利的折中。
• 经验结果表明，随着有标签数据增多，性能提升，HEAL 在不同标注比例下仍保持优势。
• 可视化结果证实学习到的超图捕捉了超出局部连接的高阶交互。