UTG

Unifying Temporal Graph (UTG) comparison between Continuous Time Dynamic Graphs and Discrete Time Dynamic Graphs (merged from UTG_dis, 3/4/2024)

Installation

1. install torch

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

2. install PyG

pip install torch_geometric==2.4.0
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.2.0+cu121.html

3. install TGX dependencies

see the instruction here

4. install wandb

pip install wandb

create branches and development

git fetch origin

git checkout -b [branch] origin/[branch]

Workflow

1. DTDG Datasets (to compare CTDG with DTDG)

a. find a discretization level that has no time gap (no empty snapshots in between)
b. generate the negative samples for val and test with `data_script.main_dtdg_gen_ns.py`
c. run experiments with any script starting with `dtdg_` in the root

2. TGB Datasets (to compare CTDG with DTDG)

a. always continuous and negative samples are downloaded from TGB
b. run experiments with any script starting with `ctdg_` in the root. 

3. run CTDG methods with discrete edge timestamps (training edges only)

a. generate the discrete timestamps with `data_script.discretize_ctdg_edges.py` will save as `.ts` files
b. load the `.ts` files to remap edge timestamps during training, see `tgn_dtdg_training.py`
c. evaluation is same as TGB

Get Started with DTDG TGX Datasets

use --wandb to turn of tracking with wandb

1. TGN commands

python dtdg_tgn.py -d canparl -t biyearly --lr 0.001 --max_epoch 500 --seed 1 --num_runs 5 --patience 100

python dtdg_tgn.py -d enron -t monthly --lr 0.001 --max_epoch 500 --seed 1 --num_runs 5 --patience 100

python dtdg_tgn.py -d uci -t weekly --lr 0.001 --max_epoch 500 --seed 1 --num_runs 5 --patience 100

python dtdg_tgn.py -d mooc -t daily --lr 0.001 --max_epoch 500 --seed 1 --num_runs 5 --patience 100

python dtdg_tgn.py -d social_evo -t daily --lr 0.001 --max_epoch 500 --seed 1 --num_runs 5 --patience 100

python dtdg_tgn.py -d contacts -t hourly --lr 0.0001 --max_epoch 200 --seed 1 --num_runs 5 --patience 50

2. HTGN commands

python dtdg_main_htgn.py --model=HTGN --dataset=canparl -t biyearly --lr 0.001 --max_epoch 500 --num_runs 5 --patience 100

python dtdg_main_htgn.py --model=HTGN --dataset=enron -t monthly --lr 0.001 --max_epoch 500 --num_runs 5 --patience 100

python dtdg_main_htgn.py --model=HTGN --dataset=uci -t weekly --lr 0.001 --max_epoch 500 --num_runs 5 --patience 100

python dtdg_main_htgn.py --model=HTGN --dataset mooc -t daily --lr 0.001 --max_epoch 500 --num_runs 5 --patience 100

python dtdg_main_htgn.py --model=HTGN --dataset social_evo -t daily --lr 0.001 --max_epoch 500 --num_runs 5 --patience 100

python dtdg_main_htgn.py --model=HTGN --dataset contacts -t hourly --lr 0.001 --max_epoch 200 --num_runs 5 --patience 50

3. EdgeBank Commands

python dtdg_edgebank.py -d enron -t monthly --mem_mode unlimited

python dtdg_edgebank.py -d enron -t monthly --mem_mode fixed_time_window

4. GCLSTM Commands

python dtdg_gclstm.py --dataset=enron -t monthly --lr 0.001 --max_epoch 500 --num_runs 5 --patience 100

python dtdg_gclstm.py --dataset=uci -t weekly --lr 0.001 --max_epoch 500 --num_runs 5 --patience 100

python dtdg_gclstm.py --dataset mooc -t daily --lr 0.001 --max_epoch 500 --num_runs 5 --patience 100

python dtdg_gclstm.py --dataset social_evo -t daily --lr 0.001 --max_epoch 500 --num_runs 5 --patience 100

python dtdg_gclstm.py --dataset contacts -t hourly --lr 0.001 --max_epoch 200 --num_runs 5 --patience 50

python dtdg_gclstm.py --dataset=canparl -t biyearly --lr 0.001 --max_epoch 500 --num_runs 5 --patience 100

EGCNO Commands

python dtdg_egcno.py --dataset=enron -t monthly --lr 2e-4 --max_epoch 500 --num_runs 5 --patience 100

python dtdg_egcno.py --dataset=uci -t weekly --lr 2e-4 --max_epoch 500 --num_runs 5 --patience 100

python dtdg_egcno.py --dataset mooc -t daily --lr 2e-4 --max_epoch 500 --num_runs 5 --patience 100

python dtdg_egcno.py --dataset social_evo -t daily --lr 2e-4 --max_epoch 500 --num_runs 5 --patience 100

python dtdg_egcno.py --dataset contacts -t hourly --lr 2e-4 --max_epoch 200 --num_runs 5 --patience 50

python dtdg_egcno.py --dataset=canparl -t biyearly --lr 2e-4 --max_epoch 500 --num_runs 5 --patience 100

5. basic GCN encoder and MLP decoder with test update

python dtdg_utg_gcn.py --dataset=enron -t monthly --lr 2e-4 --max_epoch 500 --num_runs 5 --patience 100


python ctdg_utg_gcn.py --dataset=tgbl-wiki -t hourly --lr 2e-4 --max_epoch 500 --seed 1 --num_runs 5 --patience 50 --batch_size 200

python ctdg_utg_gcn.py --dataset=tgbl-review -t monthly --lr 2e-4 --max_epoch 200 --seed 1 --num_runs 5 --patience 20 --batch_size 1000

6. CTDG commands for dt datasets

python ctdg_egcno.py --dataset=tgbl-wiki -t hourly --lr 2e-4 --max_epoch 500 --seed 1 --num_runs 5 --patience 50 --batch_size 200

python ctdg_egcno.py --dataset=tgbl-review -t monthly --lr 2e-4 --max_epoch 200 --seed 1 --num_runs 5 --patience 20 --batch_size 200

python ctdg_gslstm.py --dataset=tgbl-wiki -t hourly --lr 2e-4 --max_epoch 500 --seed 1 --num_runs 5 --patience 50 --batch_size 200

python ctdg_gclstm.py --dataset=tgbl-review -t monthly --lr 2e-4 --max_epoch 200 --seed 1 --num_runs 5 --patience 20 --batch_size 200

Get Started with TGB Datasets

use --wandb to turn on tracking with wandb

python ctdg_main_htgn.py --model=HTGN --dataset=tgbl-wiki -t hourly --lr 0.001 --max_epoch 200 --num_runs 5 --patience 50

python ctdg_main_htgn.py --model=HTGN --dataset=tgbl-review -t monthly --lr 0.001 --max_epoch 200 --num_runs 5 --patience 50

python ctdg_main_htgn.py --model=HTGN --dataset=tgbl-coin -t hourly --lr 0.001 --max_epoch 200 --num_runs 5 --patience 50

Generate negative samples for discrete datasets

python main_dtdg_gen_ns.py -d uci -t weekly

python main_dtdg_gen_ns.py -d enron -t monthly

python main_dtdg_gen_ns.py -d mooc -t daily

python main_dtdg_gen_ns.py -d social_evo -t daily

Environment

required dependencies for normal environment

1. install TGB and TGX locally, clone the repo respectively and

pip install -e .

2. install torch, PyG and other dependencies

pip install torch==1.13.1
pip install torch-geometric==2.3.1 
pip install torch-scatter==2.1.1
pip install geoopt
pip install -r requirements.txt

for compute canada, use:

pip install -r requirements_ccai.txt

Running discretized edges for CTDG methods

1. convert CTDG edgelist into DTDG edgelist (for the training set)

2. store the converted DTDG edgelist (with converted UNIX timestamps)

3. load the DTDG training set with TGB framework (or construct separate data loading / data class)

4. train TGN on DTDG training set

5. use the TGB class for evaluation set (the test edges) and evaluation