GNN frameworks
Published:
GNN frameworks
Two main types of tasks we can perform with GNN:
- node-focused
- graph-focused
Node-focused
As a general framework to build up GNNs focused on node tasks (i.e., predict the label of a node), the GNN is a composition of one or more graph filtering and non-linear activation layers
Graph-focused
As a general framework to build up GNNs focused on graph tasks, GNN is a composition of one or more graph filtering and non-linear activation layers, followed by pooling layers. It generates sparse graphs.
The graph filtered and nonlinear activation layers have the same goal as in the node-focused tasks. They are used to generate better features for each node.
The pooling layer is focused on generating higer-level features (i.e., summarize the node features). The new features capture information about the graph as a whole.
As in the example, one pooling layer follows several filter and activation layers. This is called a block and GNNs for graph-focused tasks are usually build up as the combination of one or more of those blocks
