Heterogeneous data random split for GNN

[1AngelUS]

Hi,
I am implementing a Hetero GNN by following the following link:
https://github.com/pyg-team/pytorch_geometric/blob/master/examples/hetero/hetero_conv_dblp.py

But in my data there are some nodes that don't have any label. Could you please help me as to how to random split the data to form train and test masks and train and test sets respectively.

Currently I am splitting it this way:
n = merc_x.shape[0]
idx0=list(y0[:,0]) #label 0 samples
idx1=list(y1[:,0]) #label 1 samples

random.shuffle(idx0)
random.shuffle(idx1)

train_mask_idx1 = idx1[:4000]
train_mask_idx0 = idx0[:600000]

But the model is giving very poor performance and I think it is because of this splitting strategy.

Thanks very much!

[rusty1s]

The poor performance might be due to the imbalance of labels, it looks like you have a ratio of 1:15.

To understand your code better, how is y0 and y1 defined?

[1AngelUS]

Thanks once again for your kind reply!

I ran my model after replacing SAGEConv layers by MLP and the performance was similar to the lightGBM model as you also suggested.
What shall I do to improve GNN model performance?

I was thinking may be to train the GNN in unsupervised setting. Could you help me with some source for the same.

[rusty1s]

In that case, I assume there is something wrong with your GNN or your given graph structure. How are your nodes connected to each other? Does the same hold true when using other GNN layers such as GCNConv or GATConv?

[1AngelUS]

It is bipartite graph so only one type of edges and its a dense network.

I will surely try other GNNs as you suggested and will let you know the results. I would also like to try an unsupervised GNN. Could you please help me with code sample for that?

[rusty1s]

If you want to perform message passing in a bipartite graph, I assume you need both directions. Otherwise, one set of nodes will never receive any updates.

[1AngelUS]

Thank you for pointing this out. That could be an issue, I will change my code accordingly now.