Try to understand the gradient of testing node

[juanshu30]

In the example captum_explainability.py, I didn't quite understand the code(node level explanation):

captum_model = to_captum(model, mask_type='node', output_idx=output_idx)
ig = IntegratedGradients(captum_model)
ig_attr_node = ig.attribute(data.x.unsqueeze(0), target=target,
additional_forward_args=(data.edge_index),
internal_batch_size=1)

If the target node is a test node, then how do we get the gradient of the test node feature. I guess it should be the gradient of loss with respect to the test node feature, then I am just wondering this loss is the training loss or the loss of all the nodes(including training node, val node and testing node)?

Thanks!

[rusty1s]

ig_attr_node will return the importance of node features across all nodes and all features. If you only care about the importance of features of the test node, you can do ig_attr_node.squeeze(0)[output_idx].

[juanshu30]

ig_attr_node will return the importance of node features across all nodes and all features. If you only care about the importance of features of the test node, you can do ig_attr_node.squeeze(0)[output_idx].

Thanks for your reply. I mean when calculating the integrated gradient of a single node feature (gradient of loss w.r.t feature), the loss is the node-wise(a single node) loss or the training loss on all training nodes? Thanks!

[rusty1s]

It is based on all the training nodes that are reachable within the k-hop neighborhood around your node that you want to explain (due to the way message passing works and which nodes are able to determine the given node label).

[juanshu30]

Thanks! You mean we will not use the label of this test node when we explain it? We just use the loss calculated by its k-hop training nodes, and then take derivative of this loss with respect to the test node features?

How about edges? If an edge has two test node on it end, how it is explained?

Thanks!

[rusty1s]

For explaining, we will use the label/output of the node we want to predict (this is the target in the code snippet above). The same holds true for edges, where we have a target/output per edge. For more information about how IntegratedGradients work, please also have a look at this tutorial.