Get original labels back

[AdarshMJ]

Hi!

So I have graphs in the PyG format -
PyG graph DataBatch(x=[554, 11], edge_index=[2, 1320], edge_attr=[1320, 4], y=[31, 19], pos=[554, 3], idx=[31], name=[31], z=[554], batch=[554], ptr=[32])

I convert them to networkx graphs as follows -
to_networkx(graph , node_attrs=['x'], graph_attrs=['y']))

On these NetworkX graphs I have to perform some manipulations, like adding edge/deletions based on some criteria and save them back to PyG graphs. But on doing this, I am not able to get the original labels back, basically I would want to be able to work on the modified graph using some GNN.

After manipulating the graph if I convert them back to PyG graphs this is what I get -

Data(edge_index=[2, 1322], weight=[1322], num_nodes=554)

The original information is lost. Is there any way to get the original labels back? Any insights are appreciated.

Thank you!

[rusty1s]

Thanks for reporting. Do you have a small example to reproduce? My guess here is that your transformations on the networkx graph will remove any data attached to it. Does list(G.nodes(data=True)) still hold information about x?

[AdarshMJ]

Hi! Thank you for your response.

I don't have a small example to reproduce, but I can share the Colab file. This is the error I get for

list(G.nodes(data=True))

Im using this code - https://github.com/abojchevski/node_embedding_attack to add just a single edge to the graph, all the manipulations for this requires the graph to be in networkx/scipy_sparse matrix format. After introducing the edge I want to convert it back to PyG to train.

[rusty1s]

Can you call list(G.nodes(data=True)) on the networkx graph (after doing your modifications and before calling from_networkx) rather than on the PyG data object?

[AdarshMJ]

Hi Yes, this is the output I get -

[rusty1s]

Then this looks like a problem within your networkx routine. If the networkx graph loses track of node features, then PyG cannot add them via from_networkx as well.

[AdarshMJ]

Oh okay. I probably have to rethink about the method Im using to introduce edges. Thank you!