Summary

This chapter introduced a new type of graph with spatiotemporal information. This temporal component is helpful in many applications, mostly related to time series forecasting. We described two types of graphs that fit this description: static graphs, where features evolve over time, and dynamic graphs, where features and topology can change. Both of them are handled by PyTorch Geometric Temporal, PyG’s extension dedicated to temporal graph neural networks.

Additionally, we covered two applications of temporal GNNs. First, we implemented the EvolveGCN architecture, which uses a GRU or an LSTM network to update the GCN parameters. We applied it by revisiting web traffic forecasting, a task we encountered in Chapter 6, Introducing Graph Convolutional Networks, and achieved excellent results with a limited dataset. Secondly, we used the MPNN-LSTM architecture for epidemic forecasting. We applied to the England Covid dataset a dynamic graph with a temporal signal, but...