Why does the transductive link prediction task not exclude unseen nodes from testing?

[zhiqiuyuan]

Sorry to bother you again!

I noticed in your paper that "This task has two settings: the transductive setting aims to predict future links between nodes that are observed during training, and the inductive setting predicts future links between unseen nodes." However, in your code implementation, I observed that the test set for transductive link prediction includes both observed nodes and unseen nodes.

I also reviewed the original code implementation of TGAT and TGN and found that their test sets in the transductive link prediction task are formed in a similar way to yours (i.e., containing both observed nodes and unseen nodes).

Now I am unsure whether the test set in the transductive link prediction task should exclude or include the unseen nodes.

Could you please tell me why you choose to include the unseen nodes in the test set for the transductive link prediction task? Thank you very much!

[Chandler-Q]

It is worth noting that unobserved nodes in the test set are not used for training. In this case, the authors tested the model for transductive and inductive prediction.