来源：Understanding Convolutions on Graphs

---

中文摘要 #

本文深入探讨了图神经网络(GNN)中的核心概念——图卷积运算。图卷积是将传统卷积神经网络中的卷积操作扩展到图结构数据的关键技术。文章详细阐述了图卷积的基本构建模块和设计选择，包括如何在不规则的图结构上定义卷积操作、如何聚合邻域信息、以及不同类型的图卷积层的特点。同时讨论了图卷积网络的各种设计考虑因素，如空间局部性、参数共享、计算效率等。这些基础概念对于理解和设计更高效的图神经网络架构至关重要，对于处理社交网络、分子结构、知识图谱等图结构数据具有重要应用价值。

**关键词：**图神经网络、图卷积、深度学习、神经网络架构、图结构数据

English Summary #

Understanding Convolutions on Graphs

This article delves into the core concept of graph convolution operations in Graph Neural Networks (GNNs). Graph convolution is a key technique that extends traditional convolutional operations from regular grid-like structures to irregular graph-structured data. The article elaborates on the fundamental building blocks and design choices of graph convolutions, including how to define convolution operations on irregular graph structures, methods for aggregating neighborhood information, and characteristics of different types of graph convolutional layers. It also discusses various design considerations such as spatial locality, parameter sharing, and computational efficiency. These foundational concepts are crucial for understanding and designing more efficient graph neural network architectures, with important applications in processing social networks, molecular structures, knowledge graphs, and other graph-structured data.

**Keywords: **Graph Neural Networks, Graph Convolution, Deep Learning, Neural Architecture, Graph-structured Data