TASER: Temporal Adaptive Sampling for Fast and Accurate Dynamic Graph Representation Learning

Setup

1. Setup a Python environment (>=3.11). Install PyTorch (>=2.0.1) and Deep Graph Library (>=1.1).

2. Install nvcc for cuda compilation. Make sure to choose the compatible cuda version with your PyTorch.

       conda install cuda -c nvidia/label/cuda-11.8.0
   

3. Build temporal_sampling GPU operator

       cd src/temporal_sampling/
       python setup.py build_ext --inplace
   

Download and Preprocess Dataset

1. Download datasets to the DATA/ folder.

   Note: A preprocessed version of the datasets we used is available on Google Drive. If you use this version, you can skip the preprocessing steps below.

2. Convert edge CSV to the Temporal-CSR format

       python src/gen_graph.py --data WIKI
   

3. Preprocess negative edges

       python src/preprocess.py --data WIKI --clip_root_set
   

TASER+TGNN co-training

    python src/train.py --config config_train/tgat_wiki/TGAT.yml \
                        --data WIKI \
                        --gpu 0 \
                        --cache \
                        --cached_ratio 0.2  

Important Arguments:

• --config: Config of TASER+TGNN reported in the paper. The configs of other datasets/models are under the config_train folder.
• --data: The training datasets. Available choices [WIKI, REDDIT, Flight, MovieLens, GDELT]
• --cache: Enable GPU caching
• --cached_ratio: Ratios of node features cached in GPU.

License

TASER is MIT licensed, as found in the LICENSE file.