Heterogenous Graph Features and Choice of Model

[karan96]

Hello all,

I am pretty new to pyg and still getting accustomed to all the methods and models pyg has to offer. My task in hand is to create a heterogeneous graph with two nodes {'location', 'experts'} where locations are mapped from 1 to N for a list of countries C, and experts are mapped from 1 to M from a list of experts E where each expert needs to have an attribute of {'skills'}. A typical attribute for one such node from the experts set looks like this set where each expert is separated by a slash: - {H01L41/297B41J2/14233H01L41/0471B41J2}. I need to create a hetero graph for link prediction but I am not sure how and which model to begin with. With a little bit of study I think GRAPHSAGE should be the choice of model for me. But the tutorial on how to load the data from csv files to Heterodata class are not very good in pyg. I am not sure how each attribute of skill should like, whether it can be a string like above? Or does it need to be a row of sparse matrix converted into tensor and then use each tensor as an attribute to one of the nodes of experts set. Currently the dataset looks like this: -

data = Data()
data.num_nodes = len(loc_mapping) + len(ex_mapping)
data.edge_index = torch.cat((edge_index, rev_edge_index), dim=1)
HeteroData(
num_nodes=13697,
edge_index=[2, 1258926]
)

Number of experts I have: 13631
Number of loc I have: 66

I am not sure how to add features to expert nodes and in what type? Plus what model should I be using? I was following this tutorial
https://medium.com/stanford-cs224w/disease-gene-interactions-with-graph-neural-networks-and-graph-autoencoders-bfae2c28f5ec

and they use GAE wrapping a GCNEncoder model which gives me a constant 0.5 accuracy throughout all the epochs, and all the tensor values coming out of the model were 0 which means this is not working. I changed the GCNEncoder to GNN, using recon loss the AUC was highly fluctuating ranging from 0.10 to 0.9. I ran 20 epochs 100 times to notice the average AUC and it resulted in 0.53 which is highly low considering this model was trained just on node data without any attribute.
Here are my questions: -

1. How should I add Attributes to experts, and in which format?
2. Which model and loss function to use in this scenario?
3. How to increase model accuracy?

Any help would be greatly appreciated. I have the dataset with me in case anyone wants to reproduce.

Best,
Karan

[rusty1s]

I just pushed this colab notebook to our documentation which seems 100% related to all your questions. Please take a look and let me know if it helps you!

[rusty1s]

This looks correct and should pass the test. Doesn't it?

[karan96]

@rusty1s No this didn't work, the output I am getting is
HeteroData(
user={ node_id=[608] },
movie={
node_id=[2706],
x=[2706, 20]
},
(user, rates, movie)={
edge_index=[2, 16749],
edge_label=[384],
edge_label_index=[2, 384]
},
(movie, rev_rates, user)={ edge_index=[2, 7655] }
)
It is also missing input_id=[128].
Upon running the model further I am getting loss in negative: -
100%|██████████| 190/190 [00:12<00:00, 14.79it/s]
0%| | 0/190 [00:00<?, ?it/s]Epoch: 001, Loss: -9813.9184
100%|██████████| 190/190 [00:12<00:00, 14.88it/s]
0%| | 0/190 [00:00<?, ?it/s]Epoch: 002, Loss: -10726193.4348
100%|██████████| 190/190 [00:11<00:00, 16.01it/s]
0%| | 0/190 [00:00<?, ?it/s]Epoch: 003, Loss: -396842080.4694
100%|██████████| 190/190 [00:11<00:00, 16.73it/s]
0%| | 0/190 [00:00<?, ?it/s]Epoch: 004, Loss: -4201236530.0131
100%|██████████| 190/190 [00:11<00:00, 15.91it/s]
Epoch: 005, Loss: -22461009673.3193

The training code I used is as follows: -
sampled_data.to(device)
pred = model.forward(sampled_data)
loss = F.binary_cross_entropy_with_logits(pred, sampled_data["user", "rates", "movie"].edge_label)

[rusty1s]

Can you install PyG from master (as outlined in the README.md)? AFAIK, the notebook expects PyG from master right now.

[karan96]

@rusty1s I did try to install the PyG as per README.md using the command conda install pyg -c pyg on a new environment with a fresh new installation of torch==1.12.1 but it still didn't solve the above two issues of: - Assertion Error: assert sampled_data["user", "rates", "movie"].edge_label.max() == 1 and loss being highly negative.

[rusty1s]

For installing from master, you need to follow these instructions. Conda will use the latest release.