For a batch, I am taking a sliding window of all articles mentioned within specific time intervals with significant overlap. For example, one graph/label pair is made out of [0,5], the next out of [1,6], and [2,7], and so on.
I think it’s wasteful to generate a ton of duplicate nodes, but also the logic for reusing the nodes seems complex. What’s the best way to train over batches of graphs that have very similar structures?