search

Xi-yuanWang / GLASS

GLASS: GNN with Labeling Tricks for Subgraph Representation Learning
30 stars 6 forks source link

SSL Task #4

Closed JhuoW closed 2 years ago

JhuoW commented 2 years ago

Hi,

In the paper, you propose to use node-level, edge-level, and subgraph-level SSL tasks to pre-train the node features. In GNNEmb.py, I can only find the link prediction task. Does it mean edge-level SSL plays a more important role than the other two SSL?

Besides, in the link prediction, in this line, you use all positive edges (i.e., edge_index of the original graph) to train the node embeddings, and then use a part of positive+negative edges as a validation set. Thus, the training edge set and validation edge set may be overlapped, right?

Thank you

Xi-yuanWang commented 2 years ago

Hi,

Sorry for my late reply.

"Does it mean edge-level SSL plays a more important role than the other two SSL?" Yes. We only use edge-level SSL in our experiments.

"The training edge set and validation edge set may be overlapped, right?" Yes. As the task is pretraining rather than link prediction, data leakage prevention is not necessary.

JhuoW commented 2 years ago

Thanks for your explanation.