This is a PyTorch implementation of the paper.

This implementation is in Python 3.6.6. Required packages are:

numpy      1.15.1
pytorch    1.0.1

To train and evaluate SAttLE with TwoMult decoding for link prediction run the main.py script as follows:

For FB15k-237

python main.py -d fb15k-237 --gpu 0 --batch-size 2048 --evaluate-every 250 --n-epochs 1500 --lr 0.001 --lr-decay 0.995 --lr-step-decay 2 --n-layers 1 --d-embed 100 --num-head 64 --d-k 32 --d-v 50 --d-model 100 --d-inner 2048 --start-test-at 1000 --save-epochs 1100 --dr-enc 0.4 --dr-pff 0.2 --dr-sdp 0.1 --dr-mha 0.3 --decoder twomult

For WN18RR

python main.py -d wn18rr --gpu 0 --batch-size 1024 --evaluate-every 250 --n-epochs 4500 --lr 0.001 --lr-decay 1 --lr-step-decay 2 --n-layers 1 --d-embed 100 --num-head 64 --d-k 32 --d-v 50 --d-model 100 --d-inner 100 --start-test-at 5010 --save-epochs 6000 --dr-enc 0.3 --dr-pff 0.4 --dr-sdp 0.1 --dr-mha 0.4 --decoder twomult

To train with tucker decoder change "decoder" argument to "tucker" and set the other hyper-parameters based on the paper.

SAttLE results with TwoMult decoding

SAttLE results with Tucker decoding

Please cite the following paper if you find it useful:

@article{BAGHERSHAHI2022110124,
title = {Self-attention presents low-dimensional knowledge graph embeddings for link prediction},
journal = {Knowledge-Based Systems},
pages = {110124},
year = {2022},
issn = {0950-7051},
doi = {https://doi.org/10.1016/j.knosys.2022.110124},
author = {Peyman Baghershahi and Reshad Hosseini and Hadi Moradi},
}