Heterogeneous graph classification

[Zytrus]

Hello!

I was wondering how do you know, which model you're supposed to use for your current dataset?
My dataset consists of 3 types, let's call them A, B and C and 5 relation types. Now I want to make a graph classification, so I'm wondering out of the existing models and self made ones, when do you make one yourself and when do you just take an existing one?

Are there some kind of hints that can direct you to certain models? Like what kind of feature you have (in my case node degrees) or how large the dataset is?

When would it be smart to customize your own GNN? When is it more favorable to use an existing one?

Best regards, Zytrus

[rusty1s]

This is super hard to tell and is mostly based on experimental evaluation. In general, it is good to have a basic understanding of the dataset and which features might be important. This helps you to determine, e.g. the choice of aggregation function (mean, sum, max). For example, if you only have node degrees as features, you are solely able to learn from the graph structure, and so you would want to use layers specially designed for this task with the necessary expressive power (such as GINConv). Designing a new layer is certainly a more research centric approach. One usually starts with a baseline and tries to find out why it may underperform or how results can be improved.

[rusty1s]

You can access the metadata via dataset[0].metadata(). For the number of input features across different node and edge types, we recommend that you make use of lazy initialization to simplify this process, i.e. by just passing -1 as in_channels.

[Zytrus]

But what if I do need to access the num_features? When I go through the variables step by step, the num_classes is initialized, but num_features only show an AttributeError.... with no clear specification. I tried to follow back to the dictionary _store.. but still couldn't find the cause. Is it supposed to initialize itself?

[rusty1s]

The num_features call will fail for heterogeneous graphs since this may very across node types. As a simple workaround, you can receive it for each node type individually:

for node_type in dataset[0].node_types:
    print(node_type, data[node_type].num_features)

[Zytrus]

Alright, thank you for the replies!

[rusty1s]

I integrated this as part of HeteroData in #4504. You should now be able to do:

print(hetero_data.num_node_features)
print(hetero_data.num_edge_features)