ibalazevic/multirelational-poincare

Multi-relational Poincaré Graph Embeddings

Multi-relational link prediction in the Poincaré ball model of hyperbolic space.

This codebase contains PyTorch implementation of the paper:

Multi-relational Poincaré Graph Embeddings. Ivana Balažević, Carl Allen, and Timothy M. Hospedales. Neural Information Processing Systems (NeurIPS), 2019. [Paper]

Link Prediction Results

Running a model

To run the model, execute the following command:

 CUDA_VISIBLE_DEVICES=0 python main.py --model poincare --dataset WN18RR --num_iterations 500 
                                       --nneg 50 --batch_size 128 --lr 50 --dim 40 

Available datasets are:

FB15k-237
WN18RR

To reproduce the results from the paper, use learning rate 50 for WN18RR and learning rate 10 for FB15k-237.

Requirements

The codebase is implemented in Python 3.6.6. Required packages are:

numpy      1.15.1
pytorch    1.0.1

Citation

If you found this codebase useful, please cite:

@inproceedings{balazevic2019multi,
title={Multi-relational Poincar$\backslash$'e Graph Embeddings},
author={Bala{\v{z}}evi{\'c}, Ivana and Allen, Carl and Hospedales, Timothy},
booktitle={Advances in Neural Information Processing Systems},
year={2019}
}