Forecasting web traffic

In this section, we will predict the traffic of Wikipedia articles (as an example of a static graph with a temporal signal) using a temporal GNN. This regression task has already been covered in Chapter 6, Introducing Graph Convolutional Networks. However, in that version of the task, we performed traffic forecasting using a static dataset without a temporal signal: our model did not have any information about previous instances. This is an issue because it could not understand whether the traffic was currently increasing or decreasing, for example. We can now improve this model to include information about past instances.

We will first introduce the temporal GNN architecture with its two variants and then implement it using PyTorch Geometric Temporal.

Introducing EvolveGCN

For this task, we will use the EvolveGCN architecture. Introduced by Pareja et al. [1] in 2019, it proposes a natural combination of GNNs and Recurrent Neural Networks (RNNs). Previous...