Explainer for bipartite graphs

[paulofpimenta]

Hi,

I have a graph loaded from a custom (but simple), dataset that looks like the MovieLensSmall dataset, but with many more features. Its is also a bipartite graph and I successfully loaded the graph using the load_hereto_csv example code. The dataset is about sources of information (node) that declares (edge) a disease outbreak event (node). My objective is to explain what are source or event features that contributes the most for a declaration (edge) of an event to happen. I did also some one hot encoding on categorical variables before loading the data

I have tried almost all explainers in the examples folder. The best result I have is the GNNExplainer, which fails on explanation.visualize_feature_importance. Investigating a bit further, the function fails because explainer.get('node_mask') is not empty, but a tensor of NA'S (which triggers a matplotlib fail on AxisX contaning NA values)

I do not know if I am using the right approach here. Any answer helping clarifying the issue would be highly appreciated.

Thanks in advance,

Paulo

[rusty1s]

Can you show me the error log. What do you mean by node_mask containing N/As?

[rusty1s]

Mh, can you try to add some node features to users, e.g., data['source'].x = torch.eye(212)?

[paulofpimenta]

Hello,

I tried the following (before the link-level split) :

# Add user node features for message passing:
data['source'].x = torch.eye(data['source'].num_nodes, device=device)
del data['source'].num_nodes

# Add a reverse ('movie', 'rev_rates', 'user') relation for message passing:
data = T.ToUndirected()(data)
data['source', 'event'].edge_label = data['source',
                                        'event'].edge_label.to(torch.float)
del data['event', 'rev_declares', 'source'].edge_label  # Remove "reverse" label.

but I still get the plot like the one above. Is there another way to extract the score from the explainer ? When I print my explanation defined like:

explanation = explainer(
    data.x_dict,
    data.edge_index_dict,
    index=index,
    edge_label_index=data['source', 'event'].
    edge_label_index,
) 

I get this result:

Generated explanations in ['edge_mask', 'node_mask']

HeteroExplanation(
  prediction=[777],
  target=[777],
  index=[2],
  edge_label_index=[2, 777],
  event={
    node_mask=[215, 62],
    x=[215, 62],
  },
  source={
    node_mask=[212, 212],
    x=[212, 212],
  },
  (source, declares, event)={
    edge_mask=[777],
    edge_index=[2, 777],
  },
  (event, rev_declares, source)={
    edge_mask=[777],
    edge_index=[2, 777],
  }
)

Any ideas ? Thank you :-)

[rusty1s]

Ok, so the issue is in visualization? node_mask and edge_mask seems to be well-defined in the HeteroExplanation plot. Do they actually contain values larger than zero?

[paulofpimenta]

Hello again,

Well, now I not sure that the issue is the visualization, since I get many zeros. Since I have many categorical values, I did one-hot encoding on my dataframe before applying the IdentityEncoder present in the load csv example. The result of explanation.get_node_store("source") is:

{'node_mask': tensor([[nan, nan, nan,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        ...,
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.]], device='cuda:0', dtype=torch.float64), 'x': tensor([[1., 0., 0.,  ..., 0., 0., 0.],
        [0., 1., 0.,  ..., 0., 0., 0.],
        [0., 0., 1.,  ..., 0., 0., 0.],
        ...,
        [0., 0., 0.,  ..., 1., 0., 0.],
        [0., 0., 0.,  ..., 0., 1., 0.],
        [0., 0., 0.,  ..., 0., 0., 1.]], device='cuda:0')}

On the second node (event), I have the following the I do explanation.get_node_store("event")

{'node_mask': tensor([[nan, nan, nan,  ..., nan, nan, nan],
        [nan, nan, nan,  ..., nan, nan, nan],
        [nan, nan, nan,  ..., nan, nan, nan],
        ...,
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.]], device='cuda:0', dtype=torch.float64), 'x': tensor([[1.2797e+04, 2.0180e+03, 8.0000e+01,  ..., 6.4782e+01, 6.4780e+01,
         0.0000e+00],
        [1.2797e+04, 2.0180e+03, 8.0000e+01,  ..., 6.4782e+01, 6.4780e+01,
         0.0000e+00],
        [5.0400e+02, 2.0180e+03, 8.0000e+01,  ..., 6.4782e+01, 6.4780e+01,
         2.0000e+00],
        ...,
        [1.1200e+03, 2.0190e+03, 5.0000e+01,  ..., 6.8662e+01, 2.5070e+01,
         1.6000e+01],
        [5.1500e+03, 2.0190e+03, 5.0000e+01,  ..., 6.8662e+01, 2.5070e+01,
         2.8000e+01],
        [1.1200e+03, 2.0190e+03, 5.0000e+01,  ..., 6.8662e+01, 2.5070e+01,
         1.6000e+01]], device='cuda:0')}

Finally, when I get the edge_mask with explanation.edge_items()[0][1] (maybe there is better way to do this) I get the result below:

{'edge_mask': tensor([nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
        nan, nan, nan, nan, nan, nan, nan, nan, 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
...
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0.], device='cuda:0',
       dtype=torch.float64, grad_fn=<ViewBackward0>), 'edge_index': tensor([[ 36,  67,  90,  ...,  31, 171,  13],
        [ 79, 147,  20,  ...,  52, 125,  26]], device='cuda:0')}

I can provide a colab if necessary. Thanks again for your time

Cheers,

Paulo

[rusty1s]

Wondering why you have nans in there. Does your input features contain nans as well? A Colab would be helpful.