mims-harvard/gnndelete

General Strategy for Unlearning in Graph Neural Networks

gnndelete's Issues

Sequential Unlearning

I have trained a GCN model on the Cora dataset. For the initial testing of the edge deletion task, I deleted 285 edges of training data and trained the deletion neural network.

If I further want to delete more edges then how to proceed? Do I have to again train the deletion neural network for Sequential Unlearning? If yes then why don't we directly retrain the original model instead of training the del operator again and again!!

@chengjiali @marinkaz

Problems with MIA

The official implementation of MIA is inductive learning, but GCN is often transductive learning, and the training method is different, how to judge whether it is forgotten? Can you provide the code for MIA? Thank you!

execution failures in deletion tasks

Hi there,

Upon running the command python delete_gnn.py, I encountered the following error:

  File "/home/user/Documents/code/GNNDelete/delete_gnn.py", line 86, in main
    assert args.df != 'none'
AssertionError

Similarly, when I executed the command python delete_node.py, and python delete_node_feature.py, I was met with another error:

AttributeError: 'GlobalStorage' object has no attribute 'train_pos_edge_index'

Would you mind providing some suggestions?

File Missing Issue

There are 3 prevailing errors in these codes

1. from train_mi import MLPAttacker( File may be missing)
2. from .trainer.graph_editor import GraphEditorTrainer( same)
3. from .s3io import read_txt_s3, scipy_loadmat_s3
   All modules not found error

Cannot find requirements.txt

The readme mentions:

Please install the required packages by running
pip install -r requirements.txt

Please tell where is the requirements.txt located?