Directed edges tend to consist of more information than undirected edges in the original GNN. With directed edges, we can distinguish between head and tail entity (parent and child), which allows us to perform information propagation differently for the parent and child entities.
Opposite of homogeneous graph. The heterogeneous graph contains different kinds of nodes. The simplest way to process heterogeneous graph is to convert the type of each node into one-hot vector and concatenate it to the original feature.
GraphInception – introduces metapath into the propagation of graph. Metapath groups neighbouring nodes together according to the node types and distances. For each group, GraphInception treats them as a sub-graph in homogeneous graph
Heterogeneous Graph Attention Network (HAN) utilises node-level and semantic-level attentions
Utilises edge information such as weight or type or properties of edges.
Dynamic graph has a static graph structure and dynamic input signals. To process both information, we need to collect the spatial information first, then feed the outputs into a sequence model.