Using Captum with Heterogeneous graph?

[AdrianBZG]

Hi, is there any way or "hack" to use Captum for interpretability in heterogeneous graphs? There are some resources for homogeneous but I'm unsure on how to approach this for the hetero case.

[rusty1s]

The only way I can think of is to make use of RGCNConv or RGATConv (which operate on a "typed" homogeneous graph with edge_type vector). The formulation of heterogeneous graphs as a set of bipartite graphs currently does not support explainability. I am sorry.

[AdrianBZG]

@rusty1s Thanks. How can different node types with a RGATConv approach? I currently have 2 node types in my HeteroData, one with features and labels (this one is the set of nodes to predict) and the other with no features and no labels.

[rusty1s]

You need to create some features for your second node type (random initialization or mean aggregation of neighbors usually works fine). Then, you can just concatenate the node features of the two types together.

[AdrianBZG]

@rusty1s I see, do you have an example for this? I'm not quite sure how the concatenation should be done. Thanks.

[rusty1s]

If you have a heterogeneous data object HeteroData with two node types:

data['user'].x = ...  # [num_nodes, 128]
data['paper'].x = None

you can do

data['paper'].x = torch.randn(num_paper_nodes, 128)

and call to_homogeneous() afterwards:

data = data.to_homogeneous()

This will do the concatenation internally.