Recently, graph neural networks (GNNs) have revolutionized the field of graph representation learning through effectively learned node embeddings, and achieved state-of-the-art results in tasks such as node classification and link prediction. However, current GNN methods are inherently flat and do not learn hierarchical representations of graphs---a limitation that is especially problematic for the task of graph classification, where the goal is to predict the label associated with an entire graph. Here we propose DiffPool, a differentiable graph pooling module that can generate hierarchical representations of graphs and can be combined with various graph neural network architectures in an end-to-end fashion. DiffPool learns a differentiable soft cluster assignment for nodes at each layer of a deep GNN, mapping nodes to a set of clusters, which then form the coarsened input for the next GNN layer. Our experimental results show that combining existing GNN methods with DiffPool yields an average improvement of 5-10% accuracy on graph classification benchmarks, compared to all existing pooling approaches, achieving a new state-of-the-art on four out of five benchmark data sets.
核心贡献 · Key contributions
提出 DiffPool，一种用于层次化图表示学习的可微图池化模块。 Proposes DiffPool, a differentiable graph pooling module for hierarchical graph representation learning.
实现带层次池化的深度 GNN 端到端训练，将图分类准确率提升 5-10%。 Enables end-to-end training of deep GNNs with hierarchical pooling, improving graph classification accuracy by 5-10%.
在五个基准图分类数据集中的四个上达到最先进结果。 Achieves state-of-the-art results on four out of five benchmark graph classification datasets.
学习与图中社区结构对应的可解释层次化聚类。 Learns interpretable hierarchical clusters corresponding to community structures in graphs.
证明池化操作的置换不变性，确保对节点排序的鲁棒性。 Demonstrates permutation invariance of the pooling operation, ensuring robustness to node ordering.
引入辅助链接预测和熵正则化以提高训练稳定性和聚类质量。 Introduces auxiliary link prediction and entropy regularization to improve training stability and cluster quality.
局限 · Limitations
需要预定义每层的最大聚类数，可能并非对所有图最优。 Requires predefining the maximum number of clusters per layer, which may not be optimal for all graphs.
软聚类分配可能导致粗化图噪声增加和计算成本上升。 Soft cluster assignments may lead to noisy coarsened graphs and increased computational cost.
性能依赖于基础 GNN 架构；并非所有 GNN 都能从 DiffPool 中同等受益。 Performance depends on the base GNN architecture; not all GNNs benefit equally from DiffPool.
辅助目标（链接预测、熵）增加训练复杂度，可能需要仔细调参。 Auxiliary objectives (link prediction, entropy) add training complexity and may require careful tuning.
在包含许多结构相似节点的图上表现不佳，可能将远距离节点池化在一起。 Struggles with graphs containing many structurally similar nodes, potentially pooling distant nodes together.
论文章节 · Sections（共 13）
摘要Abstract
1 引言1 Introduction
2 相关工作2 Related Work
3 提出方法3 Proposed Method
3.1 预备知识3.1 Preliminaries
3.2 基于学习分配的可微池化3.2 Differentiable Pooling via Learned Assignments
3.3 辅助链接预测目标与熵正则化3.3 Auxiliary Link Prediction Objective and Entropy Regularization
4 实验4 Experiments
4.1 基线方法4.1 Baseline Methods
4.2 图分类结果4.2 Results for Graph Classification
4.3 DiffPool 中聚类分配分析4.3 Analysis of Cluster Assignment in DiffPool