Could you explain more about how to do the link prediction evaluation in the paper?

[skx300]

Could you explain more about how to do the link prediction evaluation (predict links for t + 1) in the paper?

I am very confused about using Logistic Regression (LR) after obtaining vertex embeddings. Because you were trying to compare DeepWalk, nod2vec and TNE methods. Does it mean for every method, you will train different LR?

[luckiezhou]

Embedding methods serve as feature generators for the link prediction task (i.e. classification on whether a link will emerge or not).
To build the feature vector for a specific edge, we calculate the absolute difference between the embedding vectors of the two end nodes.

Does it mean for every method, you will train different LR?

Yes, since the feature vectors are different, we have to train a different LR for each node embedding method.