Open towardsagi opened 8 months ago
dear authors,
Usually, in DGNN, we predict links of the next snapshot using the current snapshot (and history snapshots). However, it seems that you are predicting the links using the graph contains these links.
edge_label and edge_label_index is from the current snapshot: https://github.com/pursuecong/WinGNN/blob/ad397078f59f7c071ea2fc13d3973195a05e2edb/main.py#L138-L158
and the mrr is calculated using graph_test[idx + 1] not graph_test[idx + 2] https://github.com/pursuecong/WinGNN/blob/ad397078f59f7c071ea2fc13d3973195a05e2edb/test_new.py#L46-L57
I believe the input should not play the role of output. why you use the snapshot to predict itself? Is there any misunderstanding？
@pursuecong @zhuyf8899
towardsagi
commented
8 months ago
dear authors,
Usually, in DGNN, we predict links of the next snapshot using the current snapshot (and history snapshots). However, it seems that you are predicting the links using the graph contains these links.
edge_label and edge_label_index is from the current snapshot: https://github.com/pursuecong/WinGNN/blob/ad397078f59f7c071ea2fc13d3973195a05e2edb/main.py#L138-L158
and the mrr is calculated using graph_test[idx + 1] not graph_test[idx + 2] https://github.com/pursuecong/WinGNN/blob/ad397078f59f7c071ea2fc13d3973195a05e2edb/test_new.py#L46-L57
I believe the input should not play the role of output. why you use the snapshot to predict itself? Is there any misunderstanding？