GNNExplainer improvements
#6065
Conversation
Codecov Report
@@ Coverage Diff @@
## master #6065 +/- ##
==========================================
+ Coverage 84.28% 84.30% +0.02%
==========================================
Files 362 362
Lines 20471 20444 -27
==========================================
- Hits 17254 17236 -18
+ Misses 3217 3208 -9
📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
| subgraph_edge_mask = subgraph[3][non_loop_mask] | ||
| targets.append(data.edge_label[subgraph_edge_mask].cpu()) | ||
| preds.append(expl_edge_mask[subgraph_edge_mask].cpu()) | ||
| _, _, _, hard_edge_mask = k_hop_subgraph(node_index, num_hops=3, |
There was a problem hiding this comment.
I would assume that the edges between the 3-hop nodes and the 4-hop nodes have some importance, due to the GCN normalization, right?
There was a problem hiding this comment.
4-hop nodes do not have any importance as the layer is only 3 layer deep.
There was a problem hiding this comment.
But isn't the degree calculated with the information on the 4 hop nodes? And, therefore, the output would change if we would remove all the 4 hop nodes?
There was a problem hiding this comment.
yes, but the 4 hop nodes do not have any importance score as these edge masks are never trained.
subgraphcomputation.GNNExplaineris now always applied on the full set of edges. This fixes several bugs, e.g., thatindexnow references invalid entries. Instead, we compute the hard mask to get all edges involved in message passing.GNNExplainernow correctly works with a tensor ofindextarget_indexclean-up.model.eval()to the base class and resets it state afterwards.