Comprehensive Analysis of Knowledge Graph Embedding Techniques Benchmarked on Link Prediction

Install requirements

All the code has been tested with python=3.6.10

pip install -r requirements.txt

Install the correct pytorch and torch geometric versions (i.e., the ones compatible with your cuda device). For instance:

pip install torch==1.10.0+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
pip install torch-scatter -f https://data.pyg.org/whl/torch-1.10.0+cu113.html
pip install torch-sparse -f https://data.pyg.org/whl/torch-1.10.0+cu113.html
pip install torch-geometric

Train and test R-GCN on OGB-BioKG

python train_ogb.py --wanb_log

Train and test QuatE and DualE

python pykg2vec.py <model_name> <dataset_name>

Install requirements for PyKeen

All the code has been tested with python=3.8.13

Install the correct pytorch version (i.e., the ones compatible with your cuda device). For instance:

conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge

Install PyKeen, with the wandb and plotting extension:

pip install pykeen[wandb, plotting]

N.B. For other options see https://pykeen.readthedocs.io/en/stable/installation.html

N.B. Before using wandb it is necessary to login.

For testing on ogb

conda install -c conda-forge ogb

Train and test TransE, RotatE, DistMult, ComplEx, ConvE, ConvKB, CompGCN and NodePiece

python pykeen.py <model_name> <dataset_name>

License

This project is released under the CC-BY-NC-SA 4.0 license (see LICENSE).