NBFNet: Neural Bellman-Ford Networks

PyG re-implementation of NBFNet. Authored by Zhaocheng Zhu and Michael Galkin.


NBFNet is a graph neural network framework inspired by traditional path-based methods. It enjoys the advantages of both traditional path-based methods and modern graph neural networks, including generalization in the inductive setting, interpretability, high model capacity and scalability. This repo implements NBFNet for transductive and inductive knowledge graph reasoning.


This codebase is based on PyTorch and PyTorch-Geometric. It supports training and inference with multiple GPUs or multiple machines.


You may install the dependencies via either conda or pip. Generally, NBFNet works with Python >= 3.7 and PyTorch >= 1.8.0.

From Conda

conda install pytorch=1.8.0 cudatoolkit=11.1 pyg -c pytorch -c pyg -c conda-forge
conda install ninja easydict pyyaml -c conda-forge

From Pip

pip install torch==1.8.0+cu111 -f https://download.pytorch.org/whl/torch_stable.html
pip install torch-scatter==2.0.8 torch-sparse==0.6.12 torch-geometric -f https://data.pyg.org/whl/torch-1.8.0+cu111.html
pip install ninja easydict pyyaml


To reproduce the results of NBFNet, use the following command. Alternatively, you may use --gpus null to run NBFNet on a CPU. All the datasets will be automatically downloaded in the code.

python script/run.py -c config/inductive/wn18rr.yaml --gpus [0] --version v1

We provide the hyperparameters for each experiment in configuration files. All the configuration files can be found in config/*/*.yaml.

For experiments on inductive relation prediction, you need to additionally specify the split version with --version v1.

To run NBFNet with multiple GPUs or multiple machines, use the following commands

python -m torch.distributed.launch --nproc_per_node=4 script/run.py -c config/inductive/wn18rr.yaml --gpus [0,1,2,3]
python -m torch.distributed.launch --nnodes=4 --nproc_per_node=4 script/run.py -c config/inductive/wn18rr.yaml --gpus[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]

Visualize Interpretations on FB15k-237

Once you have models trained on FB15k237, you can visualize the path interpretations with the following line. Please replace the checkpoint with your own path.

python script/visualize.py -c config/transductive/fb15k237_visualize.yaml --checkpoint /path/to/nbfnet/experiment/model_epoch_20.pth


Here are the benchmark results of this re-implementation. All the results are obtained with 4 V100 GPUs (32GB). Note results may be slightly different if the model is trained with 1 GPU and/or a smaller batch size.

Knowledge Graph Completion

Dataset MR MRR [email protected] [email protected] [email protected]
FB15k-237 113 0.416 0.322 0.456 0.602
WN18RR 650 0.551 0.496 0.573 0.661
Dataset Training Time Test Time GPU Memory
FB15k-237 22.8 min / epoch 64 s 22.9 GiB
WN18RR 12.0 min / epoch 16 s 16.1 GiB

Inductive Relation Prediction

Dataset [email protected] (50 sample)
v1 v2 v3 v4
FB15k-237 0.821 0.948 0.957 0.959
WN18RR 0.954 0.903 0.909 0.888
Dataset v1 v2 v3 v4
Training Time Test Time GPU Memory Training Time Test Time GPU Memory Training Time Test Time GPU Memory Training Time Test Time GPU Memory
FB15k-237 2 s / epoch < 1 s 3.51 GiB 6 s / epoch < 1 s 5.02 GiB 15 s / epoch < 1 s 6.56 GiB 29 s / epoch < 1 s 8.10 GiB
WN18RR 3 s / epoch < 1 s 5.17 GiB 22 s / epoch < 1 s 11.3 GiB 3 s / epoch 1 s 18.7 GiB 7 s / epoch 1 s 6.84 GiB


If you find this codebase useful in your research, please cite the original paper.

  title={Neural bellman-ford networks: A general graph neural network framework for link prediction},
  author={Zhu, Zhaocheng and Zhang, Zuobai and Xhonneux, Louis-Pascal and Tang, Jian},
  journal={Advances in Neural Information Processing Systems},


View Github