Codes for our proposed DyFMVP method.
Download the public datasets (e.g. Wikipedia) from SNAP, and store the files in folder data/. Run the following code
to generate processed files.
python utils/preprocess_data.py --data wikipedia --bipartite 172
python utils/preprocess_data.py --data paylater --bipartite 54
The generated files are named as ml_{}.csv, ml_{}.npy, ml_{}_node.npy respectively.
Detailed statistics of four popular public CTDG datasets and one large real-world financial dataset are shown as:
pandas==2.1.4
PyYAML==6.0.1
scikit-learn==1.3.2
scipy==1.11.4
six==1.16.0
torch==1.12.0
torch-scatter==2.1.0+pt112cu113
torchaudio==0.12.0
torchdiffeq==0.2.3
torchvision==0.13.0
tqdm==4.66.1
DyFMVP(-TARO):
python train_self_supervised.py --data [DATA_NAME] --msg_src left --upd_src right --restarter_type taro --restart_prob 0 --prefix [prefix] --subset [CS_ratio] --train_ratio [RS_ratio]
Default values of --subset and train_ratio are set to 1.0 as easy split setting.
Argument --subset specifies chronological subset sampling ratio, and train_ratio specifies random sampling ratio as hard split settings.
Argument restart_prob > 0 specifies using restarter on primary model DyFMVP.
python train_supervised.py --code [CODE_GEN_BY_LP] --use_valid --train_ratio [RS/CS_ratio]
Execute the train_supervised.py script for NC task (public datasets, e.g. Wikipedia, MOOC, Reddit) based on LP pretrained model.
Default value of --train_ratio is set to 1.0 as easy setting.
Argument --train_ratio can specify RS or CS ratio according to sampling choice in codes.
Argument --code is a HASH code generated by LP to establish a one-to-one correspondence.
python train_supervised_EC.py --code [CODE_GEN_BY_LP] --use_valid --train_ratio [RS/CS_ratio]
Execute the train_supervised_EC.py script for EC task (e.g. Paylater) based on LP pretrained model.
Hyper-parameter configurations of DyFMVP and TARO are shown as:
