Skip to content

Latest commit

 

History

History
198 lines (149 loc) · 9.8 KB

README.md

File metadata and controls

198 lines (149 loc) · 9.8 KB
Raw

Experiment code for PASSLEAF

Code for the PASSLEAF paper.

The implementation is a fork of UKGE.

Environment

  • python3 (>=3.6)
  • tensorflow-gpu (==1.14)

For other dependencies, please see requirements.txt. (Some of them may be UNNECESSARY!)

Datasets

The base datasets are from UKGE.

Datasets for PASSLEAF models:

  • PPI5K (ppi5k_no_psl)
  • NL27K (nl27k_no_psl)
  • CN15K (cn15k_no_psl)

The datasets locate in the data folder.

Models

In the following script examples, we use a batch size of 512, embedding dimension of 512, and save the checkpoint every 40 epochs.

PASSLEAF Model with pool-based semi-supervised learning

  • ComplEx_m5_4: Uncertain ComplEx
  • RotatE_m5: Uncertain RotatE
  • RotatE_m3_3: Simplified Uncertain Rotate
  • UKGE_logi_m2: UKGE + pool-based semi-supervised learning

Scripts:

  • Train:

    python3 run/run.py --data [dataset] --batch_size 512 --epoch 2000 -d 512 --no_psl --models_dir  [base path to save model] -m [model name]  --semisupervised_v2 --save_freq 40

  • Test:

    python3 run/test.py --data [dataset] --batch_size 512 -d 512 --no_psl --resume_model_path [saved model directory path] -m [model name] --start [starting epoch] --to [ending epoch] --step 40

  • Example:

    python3 run/run.py --data cn15k_no_psl --batch_size 512 --epoch 3000 -d 512 --no_psl --models_dir  ./trained_model_batch512_dim_512_semisupervised_v2 -m ComplEx_m5_4  --semisupervised_v2 --save_freq 40

python3 run/test.py --data cn15k_no_psl --batch_size 512 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512_semisupervised_v2/cn15k_no_psl/ComplEx_m5_4_0930 -m ComplEx_m5_4 --start 10 --to 3000 --step 20

    ＊ the model saving path name varies according to the current date.

PASSLEAF Model without pool-based semi-supervised learning

  • ComplEx_m5_1: Uncertain ComplEx (no SS)
  • RotatE_m5_1: Uncertain RotatE (no SS)
  • RotatE_m3_1: Simplified Uncertain Rotate (no SS)
  • logi: UKGE

Scripts:

  • Train:

    Same as above but without the --semisupervised_v2 flag

    python3 run/run.py --data [dataset] --batch_size 512 --epoch 2000 -d 512 --no_psl --models_dir  [base path to save model] -m [model name] --save_freq 40

  • Test:

    Same as above.

Deterministic KG Embedding Models

TODO. Not available in the repo yet

Reference

The PASSLEAF paper. Chen, Z.-M., Yeh, M.-Y., & Kuo, T.-W. (2021). PASSLEAF: A Pool-bAsed Semi-Supervised LEArning Framework for Uncertain Knowledge Graph Embedding. Proceedings of the AAAI Conference on Artificial Intelligence, 35(5), 4019-4026.

@article{
    Chen_Yeh_Kuo_2021, 
    title={PASSLEAF: A Pool-bAsed Semi-Supervised LEArning Framework for Uncertain Knowledge Graph Embedding}, 
    volume={35}, 
    url={https://ojs.aaai.org/index.php/AAAI/article/view/16522}, 
    number={5}, 
    journal={Proceedings of the AAAI Conference on Artificial Intelligence}, 
    author={Chen, Zhu-Mu and Yeh, Mi-Yen and Kuo, Tei-Wei}, 
    year={2021}, 
    month={May}, 
    pages={4019-4026} 
}

The output files and Analyses

Outputs of run/test.py

By default, run/test.py generates the following files based on saved models per training step (checkpoints):

Validation:

  • val%s_mean_rank_accurate.csv: mean rank (TEP) and hit@K (TEP) for each specified training steps.
  • val_loss_accurate.csv: MSE (CSP) and nDCG (TEP) for each specified training steps.
  • val_detail_[epoch].csv: detailed predictions. One file per training steps.

Testing: for each best checkpoint according to the validation MSE, nDCG(linear), and Hit@20

  • test_mean_rank_accurate.csv: mean rank (TEP) and hit@K (TEP) on testing set.
  • test_loss_accurate.csv: MSE (CSP) and nDCG (TEP) on testing set.
  • test_test_only_detail_[epoch].csv: detailed predictions on testing set .
  • test_mean_rank_training_included.csv: mean rank (TEP) and hit@K (TEP) on testing set. (training set candidates INCLUDED)
  • test_loss_training_included.csv: MSE (CSP) and nDCG (TEP) on testing set. (training set candidates INCLUDED)
  • test_detail_[epoch]_training_included.csv: detailed predictions on testing set. (training set candidates INCLUDED)

Automatized analysis tool

To automatize the analysis, use the autorecord_openke+ukge.py script.

python3 autorecord_openke+ukge.py > ./records.csv

Please see the script for details about the options.

Commands for reproduction

Saved trained model

Pretrained model files are missing :(

Trying to fix that.

Evaluation commands

Simplified Uncertain RotatE + SS

    python3 run/test.py --data ppi5k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512_semisupervised_v2/ppi5k_no_psl/RotatE_m3_3_0306 -m RotatE_m3_3 --start 10 --to 2000 --step 40
    python3 run/test.py --data nl27k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512_semisupervised_v2/nl27k_no_psl/RotatE_m3_3_0304 -m RotatE_m3_3 --start 10 --to 2000 --step 40
    python3 run/test.py --data cn15k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512_semisupervised_v2/cn15k_no_psl/RotatE_m3_3_0306 -m RotatE_m3_3 --start 10 --to 3000 --step 40

UKGE logi + SS

    python3 run/test.py --data ppi5k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512_semisupervised_v2/ppi5k_no_psl/UKGE_logi_m2_0306 -m UKGE_logi_m2 --start 10 --to 2000 --step 40
    python3 run/test.py --data nl27k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512_semisupervised_v2/nl27k_no_psl/UKGE_logi_m2_0301 -m UKGE_logi_m2 --start 10 --to 2000 --step 40
    python3 run/test.py --data cn15k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512_semisupervised_v2/cn15k_no_psl/UKGE_logi_m2_0301 -m UKGE_logi_m2 --start 10 --to 3000 --step 20

Uncertain ComplEx + SS

    python3 run/test.py --data ppi5k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512_semisupervised_v2/ppi5k_no_psl/ComplEx_m5_4_0306 -m ComplEx_m5_4 --start 10 --to 2000 --step 40
    python3 run/test.py --data nl27k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512_semisupervised_v2/nl27k_no_psl/ComplEx_m5_4_0301 -m ComplEx_m5_4 --start 10 --to 2000 --step 40
    python3 run/test.py --data cn15k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512_semisupervised_v2/cn15k_no_psl/ComplEx_m5_4_0301 -m ComplEx_m5_4 --start 10 --to 3000 --step 20

Simplified Uncertain Rotate (no pool-based semi-supervised learning)

    python3 run/test.py --data ppi5k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512/ppi5k_no_psl/RotatE_m3_1_0116 -m RotatE_m3_1 --start 10 --to 2000 --step 40
    python3 run/test.py --data nl27k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512/nl27k_no_psl/RotatE_m3_1_0116 -m RotatE_m3_1 --start 10 --to 2000 --step 40
    python3 run/test.py --data cn15k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512/cn15k_no_psl/RotatE_m3_1_0116 -m RotatE_m3_1 --start 10 --to 2000 --step 40

Uncertain Complex (no pool-based semi-supervised learning)

    python3 run/test.py --data cn15k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512/cn15k_no_psl/ComplEx_m5_1_0117 -m ComplEx_m5_1 --start 10 --to 2000 --step 40
    python3 run/test.py --data nl27k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512/nl27k_no_psl/ComplEx_m5_1_0117 -m ComplEx_m5_1 --start 10 --to 2000 --step 40
    python3 run/test.py --data ppi5k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512/ppi5k_no_psl/ComplEx_m5_1_0117 -m ComplEx_m5_1 --start 10 --to 2000 --step 40

UKGE (no pool-based semi-supervised learning)

    python3 run/test.py --data cn15k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512/cn15k_no_psl/logi_0117 -m logi --start 10 --to 2000 --step 40
    python3 run/test.py --data nl27k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512/nl27k_no_psl/logi_0117 -m logi --start 10 --to 2000 --step 40
    python3 run/test.py --data ppi5k_no_psl --batch_size 512 --epoch 400 -d 512 --no_psl --resume_model_path ./trained_model_batch512_dim_512/ppi5k_no_psl/logi_0117 -m logi --start 10 --to 2000 --step 40