Robust Graph Learning

This repository contains source code for the Robust Graph Learning project.

TODO:

• Add ranking metrics, and test all baselines (static, dynamic, cont) [Recall@k, NDCG@k];
• Test reddit/wiki results, to resolve OOM issues;
• Tables reporting optimal hyperparameters for each datasets.

Task:

Our code base support:

• Dynamic Link Prediction on Temporal Graph: given several time-steps of a graph dataset, our models could predict the edges change in the future time-steps.
• Link Prediction on static graph data: given one graph, predict link changes on next graph.

Data:

• DBLP
• Enron10
• Facebook
• Wikipedia
• Reddit

Attack:

We use DeepRobust to perform most of the attacks.
DeepRobust Documentation
Attack methods integrated in our framework:

• Random Attack
• Meta Attack
• Dice Attack
• Node Attack
• Temporal Attack (our own method): this method could only apply on temporal dataset.

Baseline:

Static GNNs:

• GAE
• VGAE

Discrete Dynamics GNNs:

• EvolveGCN (EGCN-H, EGCN-O)
• Euler
• DySAT

Continuous Dynamics GNN:

• TGAT

Static Defense GNN:

• Pro-GNN

Config Hyperparameter Setting:

The following table shows the hyperparameter setting for each baseline models in our default configs. Note: we perform early stop in our infrastructure pipeline.

	max_epochs	learning rate	Others
GAE	1500	enron10: 0.0001; dblp: 0.0001; fb: 0.01
VGAE	1500	enron10: 0.0001; dblp: 0.0001; fb: 0.01	same config file as GAE but change 'encoder' to 'VGCN'
EGCN-H	1500; For wiki, reddit: 800	enron10: 0.0001; dblp: 0.005; fb: 0.005
EGCN-O	1500; For wiki, reddit: 800	enron10: 0.001; dblp: 0.001; fb: 0.001
EULER	1500	enron10: 0.02; dblp: 0.02; fb: 0.02
DySAT	For enron10, dblp, fb: 100; For wiki, reddit: 50	enron10: 0.001; dblp: 0.0005; fb: 0.0005	For enron10, dblp, fb: walk_len=20; For wiki, reddit: walk_len=5, runs=5
TGAT	50	enron10: 0.0001; dblp: 0.0001; fb: 0.001

Usage:

• Using config file to train: python3 entries/train.py --cfg ./configs/link_pred_temporal_dblp.yaml
• Overwrite config by args: python3 entries/train.py --help

train.py [-h] [--gpu GPU] [--device DEVICE] [--seed SEED] [--data_name [DATA_NAME]] [--data_dir [DATA_DIR]]     [--model_name MODEL_NAME]
                [--test_time_step TEST_TIME_STEP] [--input_feat INPUT_FEAT] [--batch_size BATCH_SIZE] [--emb_dim EMB_DIM] [--lr LR]
                [--max_epoch MAX_EPOCH] [--ptb_rate PTB_RATE] [--attack_method ATTACK_METHOD] [--stopping_steps STOPPING_STEPS]
                [--log_interval LOG_INTERVAL] [--evaluate_epoch EVALUATE_EPOCH] [--runs RUNS] [--cfg CFG] [--load LOAD] [--opts ...]
robust graph learning on dynamic graphs

optional arguments:
  -h, --help            show this help message and exit
  --gpu GPU             default is to use gpu
  --device DEVICE       device id
  --seed SEED           Random seed.
  --data_name [DATA_NAME]
                        Choose a dataset from {dblp, enron10, fb}
  --data_dir [DATA_DIR]
                        Input data path.
  --model_name MODEL_NAME
                        Model name
  --test_time_step TEST_TIME_STEP
                        number of test time steps
  --input_feat INPUT_FEAT
                        0: No pretrain, 1: Pretrain with the learned embeddings, 2: Pretrain with BERT.
  --batch_size BATCH_SIZE
                        recommendation batch size.
  --emb_dim EMB_DIM     node embedding size.
  --lr LR               Learning rate.
  --max_epoch MAX_EPOCH
                        Number of epoch.
  --ptb_rate PTB_RATE   attack rate.
  --attack_method ATTACK_METHOD
                        attack_func.
  --stopping_steps STOPPING_STEPS
                        Number of epoch for early stopping
  --log_interval LOG_INTERVAL
                        Iter interval of printing loss.
  --evaluate_epoch EVALUATE_EPOCH
                        Epoch interval of evaluation.
  --runs RUNS           Independent runs with different seeds.
  --cfg CFG             specify particular yaml configuration to use
  --load LOAD           specify saved checkpoint to evaluate
  --opts ...            Command line options to overwrite configs