TGN Example

[gao462]

The example of examples/tgn.py seems to only work for events without temporal node features (or maybe I misunderstand), but the original TGN paper supports the temporal node features.

Can I ask for a better explanation or documentation on how to put temporal node features into TGNMemory? Currently, I force to merge source and destination node features in the msg input (thus msg_dim is 2 * node_dim + edge_dim).

[rusty1s]

As far as I see, you can directly add them as inputs to gnn as well, see here, indexed based on n_id from a global node feature matrix store. Does that work in your case?

[gao462]

Thank you for the reply. For me, it is a valid way to add node features.

But I am indeed more interested in the usage of TGNMemory. In the original paper, the equation (2) describes how to embed a node event as a message vector (which requires a node message network different from edge message network), but I do not see any parameter that can create a node message network.
I know that one way to do this is has MLPs to sync the node and edge feature size, then feed the node event as a self-loop edge event, but it may conflict with real self-loop event (although not exists in my current usage). So I am wondering is there any arguments to create an independent node event message network?

[gao462]

I think directly adding node features is not the best match for me.
The TGNMemory will give the latest memory state of each node, however, in a batch, we may have different temporal node features for the same node (same node at different times have different features), thus the shape of node memory states and raw node temporal features can not fit.
I checked the TGN paper, it seems that they use only the latest raw node feature in a batch (which I think is not correct since their attn convolution design is claimed to be based on TGAT which use all temporal node features). I also checked their repo, and their implementation even doesn't support using raw node features as they described in the paper.
I think the TGN model just does not support raw node features well in their design.

[rusty1s]

It's been some time since i looked into TGN. Do you know which steps are necessary to incorporate node features as part of the memory as described in the TGN paper?

[gao462]

In current TGN implementation, update_state(self, src, dst, t, raw_msg) need to become update_state_edge(self, src, dst, t, raw_msg), and we also need a new function update_state_node(self, nids, t, raw_feat) (same thing also needed for __update_msg_store__(...)). Furthermore, TGNMemory will need a new message module self.msg_node_module other than self.msg_s_module, self.msg_d_module since the edge message and node message input size can be different.

[rusty1s]

@emalgorithm