8-Graph Neural Networks
This is Lecture 8 of the series, covering deep learning methods on graph-structured data. The biggest difference from the preceding CNN and RNN lectures is that the objects being processed here are no longer images on regular grids or text on linear sequences, but non-Euclidean data whose structure is determined by connectivity relationships -- such as social networks, traffic networks, and molecular graphs.
What This Lecture Covers
From the first few pages, the focus of this material is to first explain "why GNNs are needed," then build an application-oriented perspective.
- Graph Data Background: starting from the differences between Euclidean and non-Euclidean spatial data, it explains why conventional neural networks cannot directly handle graph structures.
- Development History: provides an overview of how graph neural networks evolved from early conceptual proposals to representative methods like Graph Convolutional Networks, GraphSAGE, and GAT.
- Related Tasks: organizes problem types at the node level, edge level, and graph level to help you understand the task taxonomy of graph learning.
- Development Trends: places methods back into real-world scenarios, including traffic prediction, epidemic prediction, drug generation, and more.
How to Study This
- If your prior experience is mainly with CNNs and RNNs, the most important thing to take from this lecture is how the change in data structure drives changes in modeling -- specifically, how adjacency relationships on graphs replace regular positional structures.
- When studying, it is recommended to first clarify basic objects like nodes, edges, neighbor aggregation, and graph-level representations before looking at specific model names; otherwise, you will easily end up just memorizing abbreviations.
- This lecture works best as a topical introduction. After going through it, if you do plan to work on graph tasks, continuing to fill in message-passing mechanisms, graph convolution formulas, and graph learning frameworks will flow much more smoothly.