Skip to content
/ grail Public

Inductive relation prediction by subgraph reasoning, ICML'20

Notifications You must be signed in to change notification settings

kkteru/grail

Repository files navigation

GraIL - Graph Inductive Learning

This is the code necessary to run experiments on GraIL algorithm described in the ICML'20 paper Inductive relation prediction by subgraph reasoning.

Requiremetns

All the required packages can be installed by running pip install -r requirements.txt.

Inductive relation prediction experiments

All train-graph and ind-test-graph pairs of graphs can be found in the data folder. We use WN18RR_v1 as a runninng example for illustrating the steps.

GraIL

To start training a GraIL model, run the following command. python train.py -d WN18RR_v1 -e grail_wn_v1

To test GraIL run the following commands.

  • python test_auc.py -d WN18RR_v1_ind -e grail_wn_v1
  • python test_ranking.py -d WN18RR_v1_ind -e grail_wn_v1

The trained model and the logs are stored in experiments folder. Note that to ensure a fair comparison, we test all models on the same negative triplets. In order to do that in the current setup, we store the sampled negative triplets while evaluating GraIL and use these later to evaluate other baseline models.

RuleN

RuleN operates in two steps. Rules are first learned from a training graph and then applied on the test graph. Detailed instructions can be found here.

  • Learn rules: source learn_rules.sh WN18RR_v1
  • Apply rules:
    • To get AUC: source auc_apply_rules.sh WN18RR_v1 WN18RR_v1_ind num_of_samples_to_score(=1000)
    • To get ranking score: source auc_apply_rules.sh WN18RR_v1 WN18RR_v1_ind num_of_samples_to_score(=1000)

NeuralLP and Drum

We use the implementations provided by the authors of the respective papers to evaluate these models.

Transductive experiments

The full transductive datasets used in these experiments are present in the data folder.

GraIL

The training and testing protocols of GraIL remains the same.

KGE models

We use the comprehensive implementation provided by authors of RotatE. This gives state-of-the-art results on all datasets. The best configurations can be found here. To train these KGE models, navigate to the kge folder and run the commands as shown in the above reference. For example, to train TransE on FB237-15k, run the following command.

bash run.sh train TransE FB15k-237 0 0 1024 256 1000 9.0 1.0 0.00005 100000 16

This will store the trained model and the logs in a folder named experiments/kge_baselines/TransE_FB15k-237.

Ensembling instructions

Once the KGE models are trained, to get ensembling results with GraIL, navigate to the ensembling folder and run the following command. source get_ensemble_predictions.sh WN18RR TransE

To get ensenbling among different KGE models, from the ensembling folder run the following command. source get_kge_predictions.sh WN18RR TransE ComplEx

If you make use of this code or the GraIL algorithm in your work, please cite the following paper:

@article{Teru2020InductiveRP,
  title={Inductive Relation Prediction by Subgraph Reasoning.},
  author={Komal K. Teru and Etienne Denis and William L. Hamilton},
  journal={arXiv: Learning},
  year={2020}
}