Heterogeneous graph learning and explanation with custom dataset

[SamiDalhoumi]

Hi! I'm relatively new to PyG and I'm trying to build and explain an heterogeneous graph dataset using my own data. I managed to create the heterogeneous dataset and train a model but I'm stuck in graphs explanation step. I have some questions regarding heterogeneous graph learning using PyG:

• Why should we convert directed heterogeneous graphs to undirected graphs before learning graph models on them?

• I'm stuck in graph explanation step, I tried out different things (node-level, graph-level regression, etc.) but still getting error messages.
  

Attached a notebook and a dummy dataset to reproduce the example.

Thank you for help!
test_hybrid_gnn.zip

[wsad1]

Why should we convert directed heterogeneous graphs to undirected graphs before learning graph models on them?

This isn't required. Did you face some issue when you didn't do this?

I'm stuck in graph explanation step, I tried out different things (node-level, graph-level regression, etc.) but still getting error messages.

I looked at your notebook. What error did you get while doing node-level explanation. Could you paste the full stack trace here.

[SamiDalhoumi]

There is no isolated nodes in my graphs

[rusty1s]

Yes, but can you check during training whether there exists parameters with no gradients? For example, something like

loss.backward()
for param in model.parameters():
    assert param.grad is not None
    assert param.grad.abs().sum() > 0

[SamiDalhoumi]

It seems that there are parameters with no gradients! Is it a data or model issue? I just copied model structure from PyG documentation

[rusty1s]

Sorry for late reply. Do you know which node types/edge types do not have any gradients? What happens if you remove some types which may not be reachable within k-hop subgraphs, e.g., via del data[node_type]?

[SamiDalhoumi]

Below an example of the structure of one graph in my dataset. I convert graphs to undirected because it seems that heterogeneous GNNs don't work with directed datasets.

I tried to create graphs using only nodes types (S3, S4, S5), then (S6, S7, S8) and I've got the same issue for both cases.

The only heterogeneous graphs creation method that works for me is using existing heterogeneous operators, but I get parameters with no gradients: https://pytorch-geometric.readthedocs.io/en/latest/tutorial/heterogeneous.html