Unveiling Graph Neural Networks: Core Structures and Their Evolution

Introduction

At the beginning of 2019, a top‑ten technology outlook highlighted a trend directly related to this chapter: “massive‑scale graph neural network systems will endow machines with common sense.” This is not the first time GNNs have appeared on deep‑learning headlines. In 2018, DeepMind, Google Brain, MIT and the University of Edinburgh jointly presented a comprehensive overview of GNNs and their reasoning capabilities.

GNNs attract attention for two main reasons. First, graphs are a ubiquitous data structure lacking a universal neural model; GNNs can be seen as an extension of CNNs from Euclidean grids to non‑Euclidean graphs, transferring the convolution idea to the graph domain. Second, GNNs provide reasoning ability that traditional neural AI systems lack, effectively combining symbolic and sub‑symbolic approaches to embed rules and knowledge into neural networks.

The concept of graph neural networks dates back to 2005. Researchers in computer science and theoretical physics subsequently proposed various spatial‑domain and spectral‑domain GNN formulations, and in 2017 these two streams were unified. Since then, GNNs have received widespread interest as deep‑learning techniques mature.

Key Questions

What is the graph spectrum?

What is the graph Fourier transform?

What is a spectral‑domain graph convolutional network?

Analysis and Illustration

The following figures illustrate core concepts such as feature‑vector visualization and the structure of spectral graph convolutions.

References

Battaglia, P. W., Hamrick, J. B., Bapst, V., et al. “Relational inductive biases, deep learning, and graph networks.” arXiv preprint arXiv:1806.01261, 2018.

Bronstein, M. M., Bruna, J., LeCun, Y., et al. “Geometric deep learning: going beyond Euclidean data.” IEEE Signal Processing Magazine, 34(4):18–42, 2017.

Gori, M., Monfardini, G., Scarselli, F. “A new model for learning in graph domains.” Proceedings of the 2005 IEEE International Joint Conference on Neural Networks, 2005, pp. 729–734.

Defferrard, M., Bresson, X., Vandergheynst, P. “Convolutional neural networks on graphs with fast localized spectral filtering.” Advances in Neural Information Processing Systems, 2016, pp. 3844–3852.