What is the positive and negative edges?

[hkim716]

Hi, thanks for the awesome library for GNN.

When I saw Planetoid dataset,

Data(test_neg_edge_index=[2, 527], test_pos_edge_index=[2, 527], train_neg_adj_mask=[2708, 2708], train_pos_edge_index=[2, 8976], val_neg_edge_index=[2, 263], val_pos_edge_index=[2, 263], x=[2708, 1433])

What do the pos_edge and ned_edge mean? And what is the positive sampling and negative sampling?
Is the Neg_adj_mask equal to the adjacency matrix with shape of [number of nodes, number of nodes]?

[rusty1s]

I guess you are looking at the link prediction example. Here, pos_edge corresponds to a subset of the original edges in edge_index, which are just for training, validation, and testing respectively. neg_edge_index corresponds ground-truth negative edges (edges that are not part of the original edge_index).

[hkim716]

Ok, thanks for your reply. In my understanding, pos_edge is used for training, validation, and testing. Then why do we need neg_edge even though we have pos_edge information, which will be enough for training, validation and testing?

[rusty1s]

We need negative edges for training our model, since it is casted as a binary classification problem. You need negative edges to let the model learn how to distinguish positive from negative ones.