Details about the train/test data about seal

[muhanzhang]

Yes, PB_test.txt contains future positive links for testing. They share the same node set. Test set can contain node ids that don't appear in train, since the adjacency matrix size is determined by the maximum of all train indices and test indices. See line 78,79 in "util_functions.py". In SEAL, links are predicted based on their local structure patterns. Thus, even if a test link is far away from the train nodes community, it is still predictable as long as its local neighborhood can be sampled.

For the testing file, it should only contain those links that you are sure they exist (the true links in the future). The negative links will be sampled from the network.

[MengyangSun93]

Yes, PB_test.txt contains future positive links for testing. They share the same node set. Test set can contain node ids that don't appear in train, since the adjacency matrix size is determined by the maximum of all train indices and test indices. See line 78,79 in "util_functions.py". In SEAL, links are predicted based on their local structure patterns. Thus, even if a test link is far away from the train nodes community, it is still predictable as long as its local neighborhood can be sampled.

For the testing file, it should only contain those links that you are sure they exist (the true links in the future). The negative links will be sampled from the network.

I got it!
Thank you very much.