This repository contains the results from the reproducibility and benchmarking studies described in
Bringing Light Into the Dark: A Large-scale Evaluation of Knowledge Graph Embedding Models Under a Unified Framework.
Ali, M., Berrendorf, M., Hoyt, C. T., Vermue, L., Galkin, M., Sharifzadeh, S., Fischer, A., Tresp, V., & Lehmann, J. (2020).
arXiv, 2006.13365.
This repository itself is archived on Zenodo at .
In this study, we use the KGEMs reimplemented in PyKEEN and the authors' best reported hyper-parameters to make reproductions of past experiments. The experimental artifacts from the reproducibility study can be found here.
In this study, we conduct a large number of hyper-parameter optimizations to investigate the effects of certain aspects of models (training assumption, loss function, regularizer, optimizer, negative sampling strategy, HPO methodology, training strategy). The experimental artifacts from the ablation study can be found here.
We provide an additional tool to search through these configurations at ablation/search.py
, by finding configurations with optimal validation H@10 for a number of different queries. You can also run this script without full installation, as long as click
and pandas
are available.
General usage information can be obtained by python3 ablation/search.py --help
.
Moreover, here are a few examples:
- the overall best configuration
python3 ablation/search.py
- the best configuration for the dataset FB15k-237
python3 ablation/search.py --dataset fb15k237
- the best configuration for the dataset FB15k-237 using the distmult model
python3 ablation/search.py --dataset fb15k237 --model distmult
- the top-3 configuration for the dataset WN18-RR using LCWA training loop
python3 ablation/search.py --dataset wn18rr --training-loop lcwa --at-most 3
All configuration for installation of relevant code, collation of results,
and generation of charts is included in the tox.ini
configuration that
can be run with:
$ pip install tox
$ tox