LuckyLYM/Orca

Code repository for the SIGMOD2023 paper "Orca: Scalable Temporal Graph Neural Network Training with Theoretical Guarantees".

Orca: Scalable Temporal Graph Neural Network Training with Theoretical Guarantees

Dataset

6 datasets were used in this paper:

• Wikipedia: downloadable from http://snap.stanford.edu/jodie/.
• Reddit: downloadable from http://snap.stanford.edu/jodie/.
• MOOC: downloadable from http://snap.stanford.edu/jodie/.
• AskUbuntu: downloadable from http://snap.stanford.edu/data/sx-askubuntu.html.
• SuperUser: downloadable from http://snap.stanford.edu/data/sx-superuser.html.
• Wiki-Talk: downloadable from http://snap.stanford.edu/data/wiki-talk-temporal.html.

Preprocessing

If edge features or nodes features are absent, they will be replaced by a vector of zeros. Example usage:

python utils/preprocess_data.py --data wikipedia --bipartite
python uitls/preprocess_custom_data.py --data superuser

Requirements

• PyTorch 1.7.1
• Python 3.8
• Numba 0.54.1

Usage

Optional arguments:
    --data                  Dataset name
    --bs                    Batch size
    --n_degree              Number of neighbors to sample
    --n_head                Number of heads used in attention layer
    --n_epoch               Number of epochs
    --n_layer               Number of network layers
    --lr                    Learning rate
    --gpu                   GPU id
    --patience              Patience for early stopping
    --enable_random         Use random seeds
    --gradient              Disable gradient blocking
    --reuse                 Enable caching and reuse
    --budget                Cache size

    
Example usage:
    python train.py --n_epoch 50 --n_layer 2 --bs 200 -d wikipedia  --enable_random --reuse --lr 1e-4 --gpu 1
    python train.py --n_epoch 50 --n_layer 2 --bs 200 -d askubuntu  --enable_random --reuse --budget 1000 --lr 1e-7 --gpu 1