GLEN

This is the code associated with the paper "Boosting Temporal Graph Learning From Perspectives of Global and Local Structures". Our code references the benchmark DGB.

Running the experiments

Dependencies (with python >= 3.9):

conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge
conda install -c anaconda pandas
conda install -c anaconda scikit-learn
conda install -c conda-forge tqdm
conda install -c anaconda numpy
conda install -c conda-forge matplotlib
conda install pyg -c pyg

Datasets

The folder 'data' contains processed datasets. All the datasets used in the paper can be downloaded from here.

Run scripts (model training and evaluation)

Link prediction

To run GLEN on Wikipedia, Reddit, UCI, Enron, MOOC, UN Trade for link prediction task:

python train_link_prediction.py -d wikipedia --n_degree 10 --n_head 2 --drop_out 0.3 --window 2 --n_runs 5
python train_link_prediction.py -d reddit --n_degree 10 --n_head 4 --drop_out 0.3 --n_runs 5 --window 2 --n_runs 5
python train_link_prediction.py -d uci --n_degree 30 --n_head 1 --drop_out 0.5 --window 2 --n_runs 5
python train_link_prediction.py -d enron --n_degree 20 --n_head 4 --drop_out 0.1 --window 1 --n_runs 5
python train_link_prediction.py -d UNtrade --n_degree 20 --n_head 1 --drop_out 0.1 --window 1 --n_runs 5
python train_link_prediction.py -d mooc --n_degree 20 --n_head 4 --drop_out 0.1 --window 8 --n_runs 5

Apply DGB test for GLEN (DGB repository):

python dgb_test_for_GLEN.py -d wikipedia --n_degree 10 --n_head 2 --drop_out 0.3 --window 2 --neg_sample rnd --n_runs 5
python dgb_test_for_GLEN.py -d wikipedia --n_degree 10 --n_head 2 --drop_out 0.3 --window 2 --neg_sample hist_nre --n_runs 5
python dgb_test_for_GLEN.py -d wikipedia --n_degree 10 --n_head 2 --drop_out 0.3 --window 2 --neg_sample induc_nre --n_runs 5

python dgb_test_for_GLEN.py -d reddit --n_degree 10 --n_head 4 --drop_out 0.3 --window 2 --neg_sample rnd --n_runs 5
python dgb_test_for_GLEN.py -d reddit --n_degree 10 --n_head 4 --drop_out 0.3 --window 2 --neg_sample hist_nre --n_runs 5
python dgb_test_for_GLEN.py -d reddit --n_degree 10 --n_head 4 --drop_out 0.3 --window 2 --neg_sample induc_nre --n_runs 5

python dgb_test_for_GLEN.py -d uci --n_degree 30 --n_head 1 --drop_out 0.5 --window 2 --neg_sample rnd --n_runs 5
python dgb_test_for_GLEN.py -d uci --n_degree 30 --n_head 1 --drop_out 0.5 --window 2 --neg_sample hist_nre --n_runs 5
python dgb_test_for_GLEN.py -d uci --n_degree 30 --n_head 1 --drop_out 0.5 --window 2 --neg_sample induc_nre --n_runs 5

python dgb_test_for_GLEN.py -d enron --n_degree 20 --n_head 4 --drop_out 0.1 --window 1 --neg_sample rnd --n_runs 5
python dgb_test_for_GLEN.py -d enron --n_degree 20 --n_head 4 --drop_out 0.1 --window 1 --neg_sample hist_nre --n_runs 5
python dgb_test_for_GLEN.py -d enron --n_degree 20 --n_head 4 --drop_out 0.1 --window 1 --neg_sample induc_nre --n_runs 5

python dgb_test_for_GLEN.py -d UNtrade --n_degree 20 --n_head 1 --drop_out 0.1 --window 1 --neg_sample rnd --n_runs 5
python dgb_test_for_GLEN.py -d UNtrade --n_degree 20 --n_head 1 --drop_out 0.1 --window 1 --neg_sample hist_nre --n_runs 5
python dgb_test_for_GLEN.py -d UNtrade --n_degree 20 --n_head 1 --drop_out 0.1 --window 1 --neg_sample induc_nre --n_runs 5

python dgb_test_for_GLEN.py -d mooc --n_degree 20 --n_head 4 --drop_out 0.1 --window 8 --neg_sample rnd --n_runs 5
python dgb_test_for_GLEN.py -d mooc --n_degree 20 --n_head 4 --drop_out 0.1 --window 8 --neg_sample hist_nre --n_runs 5
python dgb_test_for_GLEN.py -d mooc --n_degree 20 --n_head 4 --drop_out 0.1 --window 8 --neg_sample induc_nre --n_runs 5

rnd means random negative sampling.

hist_nre means historical negative sampling.

induc_nre means inductive negative sampling.

Dynamic node classification

To run GLEN on Wikipedia, Reddit, MOOC for dynamic node classification task:

python pretrain_for_nc.py -d wikipedia --n_degree 10 --n_head 4 --drop_out 0.3 --window 1
python train_node_classifiction.py -d wikipedia --n_degree 10 --window 1 --n_runs 5

python pretrain_for_nc.py -d reddit --n_degree 10 --n_head 4 --drop_out 0.3 --window 2
python train_node_classifiction.py -d reddit --n_degree 10 --window 2 --n_runs 5

python pretrain_for_nc.py -d mooc --n_degree 10 --n_head 4 --drop_out 0.3 --window 8
python train_node_classifiction.py -d mooc --n_degree 10 --window 8 --n_runs 5