About the experiment in GPT-GNN about gpt-gnn HOT 4 OPEN

And if I use no_pretrain, I saw your code that is only consider train_data, valid_data, test_data in finetune_reddit.py. Am I right?

from gpt-gnn.

For all of the pre-training baselines (including GAE, GraphSAGE-unsuper, and our method), the setting all follows the pretrain-finetune paradigm, which means we first pre-train the model using the self-supervised task on the pre-training dataset (in your example, 0-70%), then use the pre-trained model to finetune on the training set (70-80%), with model selection using the valid set, and get generalization performance on the test set.

Yes, for no_pretrain, we don't leverage the pre-training data.

from gpt-gnn.

Very thanks for your reply! I'm still a little confused about GraphSAGE and GAE. Did you put all pre_train data into model or sample data just like your paper described?

from gpt-gnn.

Pre-training will use all the pre_train data, but we conduct mini-batch training by subgraph sampling to avoid memory issue (as the whole graph is too big for the GPU memory)

from gpt-gnn.