Could someone answer these questions please?

[ecoy]

Isn't there data leakage in this tutorial

I was just going through this Stochastic Training of GNN for Link Prediction — DGL 0.8.2post1 documentation I can see that the train_loader uses the whole graph and all the edges in the dataloader for the model to be trained. Doesn’t that mean that the model is also seeing the validation and test edges? Since these edges are not being removed?

Question on dataloader (link prediction)

Hi I have a dataloader for a link prediction task as follows: sampler = dgl.dataloading.MultiLayerFullNeighborSampler(n_layers) sampler = dgl.dataloading.as_edge_prediction_sampler( sampler, exclude='reverse_types', reverse_etypes={'listened': 'listened-by', 'listened-by': 'listened'}, negative_sampler=dgl.dataloading.negative_sampler.Uniform(10)) ....... def train_dataloader(self): return dgl.dataloading.DataLoader( …

Link prediction validation loss lower

What could be the reasons for validation loss to be lower in link prediction task with max-margin loss? I do transductive learning with negative sampling and I keep getting lower validation loss - cant understand why. Has anybody been in this situation and figured out why?