tedzhouhk
commented
2 years ago Thank you for your interest in our work!
It's not hard to achieve this. Here's the two places you should be looking at:
- root_nodes: You want each batch to contains all nodes since you want to rank the top destination nodes. If the number of nodes are too large, this might need to be done in a minibatch way.
- EdgePredictor: You need to compute the edge probabilities for all destination nodes and return the top x nodes.
Lastly, I want to point out that under the current link prediction setup, the top 100 destination nodes for any source node would be the same at the same time, as the edge probability is calculated by adding two "edge scores" from the source and destination nodes. Another setup for GNN edge prediction is to directly learn a embedding for each node pair, which might better serve your need.
moudheus
Currently TGL only provides training scripts. I would like to predict the top K most likely future links after training.
Example:
Alternatively, I would appreciate documentation on how to achieve this on my own. I believe it is doable by generating the list of candidate edges as a test set, and then extract the scores from the evaluation function.
Thank you.