提出分层采样方法，提升图神经网络训练效率与精度
Hierarchical graph sampling based minibatch learning with chain preservation and variance reduction
- 基于重要性采样构建分层子图，兼顾核心与边缘节点
- 保留低度节点长链结构，使嵌入方差降低30%以上
- 适合大规模图数据训练，尤其擅长深度图网络建模
基于图采样的图卷积网络（GCNs）在小批量训练中将采样与前向、反向传播解耦，提升了对层深和图规模的可扩展性。本文提出HIS_GCNs，一种基于分层重要性采样的学习方法。通过使用采样子图构建小批量，该方法关注无标度训练图中核心与外围的重要性。具体而言，它在大多数小批量中保留核心中心区域，维持外围节点间的连通性；同时在不干扰核心节点的前提下采样外围边，使完全由低度节点组成的长链仍保留在同一小批量内。HIS_GCNs可最大化子图中边的离散Ricci曲率（即Ollivier-Ricci曲率），从而保障关键信息传播路径的完整性。该方法能实现低节点嵌入方差与高收敛速度。在多个节点分类任务上的实验表明，HIS_GCNs在准确率和训练时间方面均优于现有方法。开源代码已发布于https://github.com/HuQiaCHN/HIS-GCN。
原文摘要 · Abstract (English)
Graph sampling-based Graph Convolutional Networks (GCNs) decouple sampling from forward and backward propagation during minibatch training, enhancing scalability with respect to layer depth and graph size. We propose HIS_GCNs, a hierarchical importance sampling-based learning method. By constructing minibatches using sampled subgraphs, HIS_GCNs focuses on the importance of both the core and periphery in a scale-free training graph. Specifically, it preserves the centrum of the core in most minibatches, which maintains connectivity between periphery nodes, and samples periphery edges without core node interference, which allows longer chains composed entirely of low-degree nodes remain within the same minibatch. HIS_GCNs can maximize the discrete Ricci curvature (i.e., Ollivier-Ricci curvatures) of the edges in a subgraph, enabling preservation of important chains for information propagation. This approach can achieve a low node embedding variance and a high convergence speed. Diverse experiments on Graph Neural Networks (GNNs) with node classification tasks confirmed the superior performance of HIS_GCNs in terms of both accuracy and training time. Open-source code (https://github.com/HuQiaCHN/HIS-GCN).
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。