search

AstraZeneca / NESS

Official implementation of "NESS: Node Embeddings from Static Subgraphs"
Apache License 2.0
15 stars 0 forks source link

Test Results #9

Closed zzzmm1 closed 1 year ago

zzzmm1 commented 1 year ago

I have a problem about the test results. As mentioned in the paper, "Aggregating the node representations learned from each subgraph to obtain a joint representation of the graph at test time." Since the split of the subgraphs is randomized, are the node representations uncertain at test time?

talipucar commented 1 year ago

Sorry, just saw this message. I am not sure what you mean by uncertain at test time. In this work, you generate subgraphs by randomly removing edges. It is a transductive setting, so all subgraphs contain all the nodes, but different subsets of edges (there is no overlap or shared edges between subgraphs). At test time, you first compute the node embeddings from each of the subgraphs, and then you aggregate them to get the joint embedding. Once you have the joint embedding, you can make link predictions.

I am closing it, but if it is still not clear, please feel free to re-open the issue.