This is the implementation of DRUM, proposed in the following paper:
DRUM: End-To-End Differentiable Rule Mining On Knowledge Graphs -- [arXive + appendix]. Ali Sadeghian, Mohammadreza Armandpour, Patrick Ding, Daisy Zhe Wang. NeurIPS 2019.
- Python 3.6
- Numpy
- Tensorflow 1.13.1
The following command starts training a dataset about family relations, and stores the experiment results in the folder exps/demo/
.
python src/main.py --datadir=datasets/family --exps_dir=exps/ --exp_name=demo
To get the best performance, use different ranks for different datasets, default value is set to 3.
To evaluate the prediction results, follow the steps below. The first two steps is preparation so that we can compute filtered ranks (see TransE for details).
We use the experiment from Quick Start as an example. Change the folder names (datasets/family, exps/dev) for other experiments.
. eval/collect_all_facts.sh datasets/family
python eval/get_truths.py datasets/family
python eval/evaluate.py --preds=exps/demo/test_predictions.txt --truths=datasets/family/truths.pckl
This code partially borrows from Neural LP.