Negative Graph in EdgeDataLoader

Questions
#1

Does the negative graph in EdgeDataLoader contain all the negative examples between the all the nodes in a graph? If not, then what are criteria for those selective negative examples?
I want my link prediction to predict new possible links between two nodes based on node data. That’s why I am focusing on negative_graph and negative_score. Is that the right way to go about it?
I am following this documentation

Information of edges and nodes for stochastic link prediction evaluation
#2

This PR provides a tutorial on how to use your link prediction model to evaluate on new links. Maybe you could take a look?

1 Like
#3

Thanks…I want to find prediction score for new links between two nodes, as in there is no link present between them in dataset but can be in future, based on node attributes. Should I create a disconnected graph with unrelated nodes and use predictor on it? Or should I look for negative examples(negative_graph) of given dataset…TIA

#4

In the most general case, yes; you can build a graph with edges connecting the node pairs you are interested in (even if they are not connected in real life), and run EdgeDataLoader on top of that.

1 Like
#5

I have followed the testing part from this PR. I am trying to further find pred_score for positive graphs from testing phase. I am trying to use label(==1)'s index to find out positive graph examples. Is that the proper way to do it?
Also, predictor gives score ranging from [106.82592010498047 , -37.0677375793457] for positive examples. Is it alright? cause as far as I know the predicted score is supposed to be in range of [0,1].
TIA

#6

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.