6 Dynamic Graphs: Spatial-Temporal GNNs
This chapter covers
- Introducing memory into your deep learning models
- Understanding the different ways to model temporal relations using GNNs
- Implementing dynamic GNNs
- Evaluating your temporal GNN models
We exist in a physical world which contains both space and time. When we look out a window, we might see birds flying through the sky or fruit falling from a tree. These objects are physically moving, they follow a trajectory in front of our eyes, and we are able to predict their future positions based on these observed trajectories. So far in this book, we have covered many of the different methods that we can use to combine deep learning with graph data, including message passing (Chapter 2), graph convolutional networks (Chapter 3), graph attention networks (Chapter 4), and generative graph neural networks (Chapter 5) . These methods include the fundamentals of graph-based deep learning but also how graph-based learning overlaps and builds on other deep learning concepts including attention and generative modeling.
So far, all of our models and data have been single snapshots in time. In practice, the world is dynamic and in constant flux. Traffic flow, weather patterns, the spread of diseases across networks of people are all examples where more information can be gained when modeled with spatio-temporal graphs instead of static graphs.