How to load heterogeneous temporal graph with the combination of HeteroData and TemporalData

[rangerhu]

In the case of https://pytorch-geometric.readthedocs.io/en/latest/notes/load_csv.html, if I want to combine the timestamp attr from rating.csv, thus making the constructing graph as a heterogeneous temporal graph, How shall I change the code in https://github.com/pyg-team/pytorch_geometric/blob/master/examples/hetero/load_csv.py to reach this goal.
PS: I noticed that those data types under TORCH_GEOMETRIC.DATA class include HeteroData() and TemporalData() which are designed for Heterogeneous graph and Temporal graph respectively, if I want to construct heterogeneous temporal graph, then how can I combine HeteroData() with and TemporalData() to load my data?
Can't be appreciated it too much if you can find time to answer my question.

[rusty1s]

I think in these scenarios that HeteroData is still the preferred choice here. Note that we are currently working on providing better support for this. For example, we just added support for this in torch-sparse (requires manual installation from source for now), and with the node-level time_attr attribute of the NeighborLoader class. With this, the NeighborLoader class performs temporal sampling that avoids leaking future information.

[rangerhu]

Thank you for your reply, sir. Is that mean in order to reach my goal, I will need to do the following work?

1. assign the timestamp value to data['user', 'rates', 'movie'].time_attr (in the case of Pyg docs MovieLens dataset Case)
2. upgrade my pytorch-sparse library to the latest version by manualling installation from source
3. Convert an existing temporal GNN(e.g. TGN) to heterogeneous temporal GNN using torch_geometric.nn.to_hetero() to embedding the heterogeneous temporal graph.
   PS: my current version of relevant dependency:
   pytorch 1.10.1
   pytorch-sparse 0.6.13
   pyg 2.0.4

[rangerhu]

From the GitHub page of pytorch-sparse , it seems the latest version of the library is 0.6.13, which means I will don't have to upgrade my pytorch-sparse library?

[rusty1s]

1. Currently, time_attr needs to be a node-level attribute. This is a current limitation, sorry about that :(
2. Temporal sampling support is not yet released. It will get integrated in a new torch-sparse release. For now, cloning the repo and running pip install -e . should work (but may take some time to compile).
3. Yes. You can also compare your model to TGN. This would be a very interesting case study.

[rangerhu]

got it, thank you,sir ~

[shairal1]

Hey , I also want something like this . Would really like to know if you figured out a way to do this :) .Thanks @rangerhu