Question about LastNeighborLoader

[Bowen-n]

I'm studying TGN recently and reading example code. I don't understand what LastNeighborLoader is doing.

It contains data neighbors, e_id, __assoc__, what do these variables mean?

[rusty1s]

LastNeighborLoader will return the last encountered neighbohrs of n_id in an ever-growing graph. The idea is to insert new temporal edges via insert, and sample encountered edges via __call__.

Here, e_id is used as a reference to the global index of the returned edge. This helps to gather the respective edge features from your dataset belonging to the sampled edges.
__assoc__ is used to map global edge indices to the local ones belonging to the sampled subgraph. For example, your original graph may contain 10,000 nodes, but your sampled subgraph does only reference 100 of them.

[Bowen-n]

n_id, edge_index, e_id = neighbor_loader(n_id)
...
z = gnn(z, last_update, edge_index, data.t[e_id], data.msg[e_id])

I still have some questions about LastNeighborLoader.
What does edge_index that LastNeighborLoader returns mean? And why it should be fed into gnn?

[rusty1s]

The edge_index will return a subgraph of all encountered edges, which will contain the last x interactions to other nodes from the nodes in n_id.

[Bowen-n]

It seems that current research on temporal graph networks are all applied on tasks like edge prediction or node classification. Is there any approach to apply temporal graph networks on graph-level task like graph classification?
For example, I have a temporal graph dataset. Each graph has a class label. And each edge in the graph has a timestamp representing the temporal information. And I want to conduct graph classification on this dataset.
How can I apply TGN on this task? In TGN, batches are formed from a single graph but in a graph classification task batches are formed from the whole graph dataset.
Is it possible to simply regard this dataset as a static graph dataset and just encode the timestamp to a kind of edge feature?

These questions are not about PyG. I'll be very glad if you could help. 😊

[rusty1s]

Indeed, there does not seem to be much research regarding temporal graph networks for graph-level tasks. I feel that TGN is not the right use-case here, and you may have multiple options here:

1. Regard this dataset as a static graph. This should already work just fine.
2. Create multiple graph instances w.r.t to different timestamps. You can utilize GNNs to obtain graph-level predictions on each of those graphs in isoloation, and input them into a GRU/LSTM afterwards.

[Bowen-n]

Thanks for your advice :)