QueuQ
commented
1 year ago Yes, since GNNs generates the representation of a node based on the information from its multi-hop neighbors, if we keep the inter-task edges, we are actually allowing the access to the subgraphs of the previous tasks. If we only use the subgraph of the current task, then keeping the inter-task edges or not have no difference at all.
This is the key difference caused by the inter-task edges, which is also described in the last paragraph of Section 2.2 in our paper.
When considering the edges between tasks, why is the input during training the intersection of the training set train_ids of all previous task nodes and the current task node? The training subgraph is the subgraph formed by all the previous tasks rather than the subgraph of the current task.