Hi, I’ve trained a model for link prediction. But I am kinda confused how to use it in actual inference. For example, I have a new node and want to predict links with the existing nodes in the graph, should I add this new node to the graph and train the model again to get the embedding or directly use the model to infer to get the embedding.
You do not need to train the model again. However, it will be better if there are known connections between this new node and the old nodes, which can be used in the forward computation during inference.
What if it is a entirely new node without any connections with old nodes, like I want to predict the how many times a new published paper will be cited
The task itself is ill-defined, as a new paper cannot get cited by old papers. Meanwhile, if the original graph has some isolated nodes, the trained model might be able to generalize to new isolated nodes more or less.
Yes, you’re right. I just realize this task is pretty wired. Do you have any idea on how to refine this task, like I want to predict the times that a new paper will be cited in the future
If you have access to the date of publication, then you can split the graph based on it.
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.