CrossE

INTRODUCTION

Original paper: Interaction Embeddings for Prediction and Explanation in Knowledge Graphs. (WSDM'2019)

DATASETS

There are three benchmark datasets used in this paper, WN18, FB15k and FB15k-237.

RUN

(parameters for FB15k, FB15k-237, WN18 are taken from the official paper)

for FB15k: python3 CrossE.py --batch 4000 --data datasets/FB15k/ --dim 300 --eval_per 20 --loss_weight 1e-6 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/save_FB15k/

for FB15k-237: python3 CrossE.py --batch 4000 --data datasets/FB15k-237-swapped/ --dim 100 --eval_per 20 --loss_weight 1e-5 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/FB15k-237-swapped/

for WN18: python3 CrossE.py --batch 2048 --data datasets/WN18/ --dim 100 --eval_per 20 --loss_weight 1e-4 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/WN18/

for DBpedia15k python3 CrossE.py --batch 4000 --data datasets/DBpedia15k/ --dim 100 --eval_per 20 --loss_weight 1e-5 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/save_DBpedia15k/

for DBpediaYAGO python3 CrossE.py --batch 1000 --data datasets/DBpediaYAGO/ --dim 300 --eval_per 20 --loss_weight 1e-5 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/save_DBpediaYAGO/

python3 CrossE.py --batch 4000 --data datasets/FB15k-237-swapped/ --dim 100 --eval_per 20 --load_model ./save/FB15k-237-swapped/CrossE_DEFAULT_499.ckpt --loss_weight 1e-5 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/FB15k-237-swapped/

Restore training

If you want to restore the training phase over a dataset you need to specify the parameter --load_model.

E.g.: if you want to restore the data from the experiment made on the reduced version of FB15K dataset --batch 20 --data datasets/FB15k/reduced1.1/ --dim 2 --eval_per 20 --loss_weight 1e-6 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/save_FB15k_reduced1.1/ --load_model ./save/save_FB15k_reduced1.1/CrossE_DEFAULT_499.ckpt

for FB15k: python3 CrossE.py --batch 4000 --data datasets/FB15k/ --dim 300 --eval_per 20 --loss_weight 1e-6 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/save_FB15k/ --load_model ./save/save_FB15k/CrossE_DEFAULT_340.ckpt

for WN18: python3 CrossE.py --batch 2048 --data datasets/WN18/ --dim 100 --eval_per 20 --loss_weight 1e-4 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/WN18/ --load_model ./save/WN18/CrossE_DEFAULT_460.ckpt

for DBpedia15k python3 CrossE.py --batch 4000 --data datasets/DBpedia15k/ --dim 100 --eval_per 20 --loss_weight 1e-5 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/save_DBpedia15k/ --load_model ./save/save_DBpedia15k/CrossE_DEFAULT_160.ckpt

for DBpediaYAGO python3 CrossE.py --batch 1000 --data datasets/DBpediaYAGO/ --dim 300 --eval_per 20 --loss_weight 1e-5 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/save_DBpediaYAGO/ --load_model ./save/save_DBpediaYAGO/CrossE_DEFAULT_180.ckpt --worker 10

Pay attention to the checkpoint file to specify: in the save folder, you'll see files like these but none of them has to be specified to restore the model; instead, you will just specify a name like CrossE_DEFAULT_499.ckpt without any other extension (TensorFlow will manage it).

#EXPLANATION PROCESS The explanation process is an extension of the original project, implemented following the algorithm of the original paper of CrossE. To execute the explanation process it is necessary to follow these steps:

1. Train the model on the desired dataset

2. Load the model from the last iteration in order to save other useful data for the further steps

   • for FB15k-237: python3 CrossE.py --batch 4000 --data datasets/FB15k-237-swapped/ --dim 100 --eval_per 20 --loss_weight 1e-5 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/FB15k-237-swapped/ --load_model ./save/FB15k-237-swapped/CrossE_DEFAULT_499.ckpt

   • for FB15k: python3 CrossE.py --batch 4000 --data datasets/FB15k/ --dim 300 --eval_per 20 --loss_weight 1e-6 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/save_FB15k/ --load_model ./save/save_FB15k/CrossE_DEFAULT_499.ckpt

   • for WN18: python3 CrossE.py --batch 2048 --data datasets/WN18/ --dim 100 --eval_per 20 --loss_weight 1e-4 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/WN18/ --load_model ./save/WN18/CrossE_DEFAULT_499.ckpt

   • for DBpedia15k python3 CrossE.py --batch 4000 --data datasets/DBpedia15k/ --dim 100 --eval_per 20 --loss_weight 1e-5 --lr 0.01 --max_iter 500 --save_per 20 --save_dir ./save/save_DBpedia15k/ --load_model ./save/save_DBpedia15k/CrossE_DEFAULT_499.ckpt

3. Run the batch_similarity.py file

   • EUCLIDIAN DISTANCE (standard)

     • for FB15k-237: python3 explanation/batch_similarity.py --data ./save/FB15k-237-swapped/out_data/pickle/
     • for WN18: python3 explanation/batch_similarity.py --data ./save/WN18/out_data/pickle/
     • for DBpedia15k python3 explanation/batch_similarity.py --data ./save/save_DBpedia15k/out_data/pickle/
     • for FB15k: python3 explanation/batch_similarity.py --data ./save/save_FB15k/out_data/pickle/

   • COSINE SIMILARITY

     • for FB15k: python3 explanation/batch_similarity.py --data ./save/save_FB15k/out_data/pickle/ --distance cosine
     • for FB15k-237: python3 explanation/batch_similarity.py --data ./save/FB15k-237-swapped/out_data/pickle/ --distance cosine
     • for DBpedia15k python3 explanation/batch_similarity.py --data ./save/save_DBpedia15k/out_data/pickle/ --distance cosine
     • for WN18: python3 explanation/batch_similarity.py --data ./save/WN18/out_data/pickle/ --distance cosine

   • SEMANTIC SIMILARITY

     • for DBpedia15k python3 explanation/batch_similarity.py --data ./save/save_DBpedia15k/out_data/pickle/ --semantic_data ./datasets/DBpedia15k/

4. Run the explanation process

   • EUCLIDEAN DISTANCE BASED (default):

     • 2% of predictions (is default):

       • for FB15k: python3 explanation/explanation.py --data ./save/save_FB15k/out_data/pickle/ --save_dir explanation/results/save_FB15k/Euclidian/2perc/

       • for FB15k-237: python3 explanation/explanation.py --data ./save/FB15k-237-swapped/out_data/pickle/ --save_dir explanation/results/FB15k-237-swapped/Euclidian/2perc/

       • for DBpedia15k python3 explanation/explanation.py --data ./save/save_DBpedia15k/out_data/pickle/ --save_dir explanation/results/save_DBpedia15k/Euclidian/2perc/

       • for WN18 python3 explanation/explanation.py --data ./save/WN18/out_data/pickle/ --save_dir explanation/results/WN18/Euclidian/2perc/

       • for DBpediaYAGO python3 explanation/explanation.py --data ./save/save_DBpediaYAGO/out_data/pickle/ --save_dir explanation/results/save_DBpediaYAGO/Euclidian/2perc/

     • 5% of predictions

       • for FB15k: python3 explanation/explanation.py --data ./save/save_FB15k/out_data/pickle/ --save_dir explanation/results/save_FB15k/Euclidian/5perc/ --predictions_perc 5

       • for FB15k-237: python3 explanation/explanation.py --data ./save/FB15k-237-swapped/out_data/pickle/ --save_dir explanation/results/FB15k-237-swapped/Euclidian/5perc/ --predictions_perc 5

       • for DBpedia15k python3 explanation/explanation.py --data ./save/save_DBpedia15k/out_data/pickle/ --save_dir explanation/results/save_DBpedia15k/Euclidian/5perc/ --predictions_perc 5

       • for WN18 python3 explanation/explanation.py --data ./save/WN18/out_data/pickle/ --save_dir explanation/results/WN18/Euclidian/5perc/ --predictions_perc 5

       • for DBpediaYAGO python3 explanation/explanation.py --data ./save/save_DBpediaYAGO/out_data/pickle/ --save_dir explanation/results/save_DBpediaYAGO/Euclidian/5perc/ --predictions_perc 5

   • COSINE SIMILARITY BASED

     • 2% of predictions (is default):

       • for FB15k: python3 explanation/explanation.py --data ./save/save_FB15k/out_data/pickle/ --save_dir explanation/results/save_FB15k/cosine/2perc/ --distance cosine

       • for FB15k-237: python3 explanation/explanation.py --data ./save/FB15k-237-swapped/out_data/pickle/ --save_dir explanation/results/FB15k-237-swapped/cosine/2perc/ --distance cosine

       • for DBpedia15k python3 explanation/explanation.py --data ./save/save_DBpedia15k/out_data/pickle/ --save_dir explanation/results/save_DBpedia15k/cosine/2perc/ --distance cosine

       • for WN18 python3 explanation/explanation.py --data ./save/WN18/out_data/pickle/ --save_dir explanation/results/WN18/cosine/2perc/ --distance cosine

       • for DBpediaYAGO python3 explanation/explanation.py --data ./save/save_DBpediaYAGO/out_data/pickle/ --save_dir explanation/results/save_DBpediaYAGO/cosine/2perc/ --distance cosine

     • 5% of predictions

       • for FB15k: python3 explanation/explanation.py --data ./save/save_FB15k/out_data/pickle/ --save_dir explanation/results/save_FB15k/cosine/5perc/ --predictions_perc 5 --distance cosine

       • for FB15k-237: python3 explanation/explanation.py --data ./save/FB15k-237-swapped/out_data/pickle/ --save_dir explanation/results/FB15k-237-swapped/cosine/5perc/ --predictions_perc 5 --distance cosine

       • for DBpedia15k python3 explanation/explanation.py --data ./save/save_DBpedia15k/out_data/pickle/ --save_dir explanation/results/save_DBpedia15k/cosine/5perc/ --predictions_perc 5 --distance cosine

       • for WN18 python3 explanation/explanation.py --data ./save/WN18/out_data/pickle/ --save_dir explanation/results/WN18/cosine/5perc/ --predictions_perc 5 --distance cosine

       • for DBpediaYAGO python3 explanation/explanation.py --data ./save/save_DBpediaYAGO/out_data/pickle/ --save_dir explanation/results/save_DBpediaYAGO/cosine/5perc/ --predictions_perc 5 --distance cosine