What is a temporal GNN?

Temporal GNN (Temporal Graph Neural Network) is a type of graph neural network that considers the temporal dynamics of a graph. Unlike traditional GNNs, which operate on static graphs, temporal GNNs model changes in the structure of the graph and/or the features of nodes and edges over time.

The main idea is to integrate temporal information to predict events, the evolution of connections, or node behavior in dynamic networks, such as social networks, financial transactions, or biological systems.

Example: if you have a social network graph, a temporal GNN can account for when and how the connections between users change to predict future interactions.

A simplified architecture example in pseudocode:

for t in time_steps:
    graph_t = get_graph_at_time(t)
    node_embeddings = GNN_layer(graph_t, previous_embeddings)
    previous_embeddings = node_embeddings

Thus, temporal GNN combines graph operations with temporal models, such as recurrent networks or attention mechanisms, to process sequences of graphs.