label leakage

[zhandand]

PyGCL/examples/GRACE_SupCon.py

Line 75 in 3e6357b

 z, _, _ = encoder_model(data) 

In the supervised contrastive learning, the nodes with the same labels are collected as the postive pairs. So in my opinion, in the constrative pretraining stage, we can't utilize all the data, otherwise the labels are leaked in the tuning stage

[zhandand]

Sorry, my mistake. It seems that you have dealt with this problem

PyGCL/GCL/models/sampler.py

Line 12 in 3e6357b

 extra_pos_mask[~data.train_mask][:, ~data.train_mask] = False