KG2ART_RepLearning

• In this experiment, 5 datasets are included. The binary forms in 'pickle' format are available in 'data' folder.

• The textual format of triples of all 5 five datasets are also available in 'src_data' folder. The 5 datasets are Codex-m, Jet_Engine, Kinship, Nations and UMLS

• Preprocessing of these datasets are required for training the models. The data source to preprocess can be found in 'src_data' folder. The preprocess function can be found in preprocess_datasets.py

• In models.py:

  • For tail prediction, the following parameter should be set as 'target = RHS'.
  • For head prediction, the following parameter should be set as 'target = LHS'.
  • For relation prediction, the following parameter should be set as 'target = rel'

• Run the following query to produce the results for those models. This query could be found in 'Experiment.ipynb' For example: %run main.py --dataset Jet_Engine --score_rel True --model RESCAL --rank 200 --learning_rate 0.1 --batch_size 200 --lmbda 0.05 --w_rel 4 --max_epochs 100

• In the query, dataset name,model and hyper parameters can be specified.

• For experiments with state-of-the-art models (RESCAL, TuckER, ComplEX, ConvE, CompGCN), wandb (https://wandb.ai) is required.

• For KG2ART model, the functions for encoding and inferences are defined in dambART.py. It can be included in the jupyter notebook by calling "import dambART.py"