search

vuptran / graph-representation-learning

Autoencoders for Link Prediction and Semi-Supervised Node Classification (DSAA 2018)
MIT License
253 stars 67 forks source link

Why is link prediction setting in this literature referred to as 'unsupervised'? #4

Closed DaehanKim closed 5 years ago

DaehanKim commented 5 years ago

Hi,

I found that your literature actually uses true link information by using masked cross entropy loss. In what sense, is this setting called 'unsupervised' link prediction?

Thanks in advance.

vuptran commented 5 years ago

You raise a good point. In the common link prediction problem, we are given a partially observed graph with known links and non-links as the starting point. These are values of the data set, similar to pixel values of images. We use the values of the network to make predictions on unobserved entries in the network or adjacency matrix without requiring additional ground truth labels. In this sense, we are doing unsupervised link prediction by using only the input values of the adjacency matrix.

But your point is valid because we still need the ground truth link labels to make progress. A true unsupervised link prediction problem is where the adjacency matrix is completely unknown and the task is to predict the adjacency matrix from scratch. So what we are doing here is not entirely unsupervised link prediction, and you are correct.