/GETD

Generalizing Tensor Decomposition for N-ary Relational Knowledge Bases

Primary LanguagePythonMIT LicenseMIT

Generalizing Tensor Decomposition for N-ary Relational Knowledge Bases

Link Prediction Results on N-ary Relational KBs

Dataset MRR Hits@10 Hits@3 Hits@1
WikiPeople-3 0.373 0.588 0.401 0.284
WikiPeople-4 0.386 0.596 0.462 0.265
JF17K-3 0.732 0.856 0.764 0.669
JF17K-4 0.810 0.913 0.844 0.755

Running a model

To run the model, execute the following command:

CUDA_VISIBLE_DEVICES=0 python main.py --dataset dataset --num_iterations 200 --batch_size batch_size 
--edim edim --rdim rdim --k k --n_i n_i --TR_ranks TR_ranks --dr dr --lr lr --input_dropout --input_d 
--hidden_dropout --hidden_d

Available datasets are:

WikiPeople-3, WikiPeople-4, JF17K-3, JF17K-4

To reproduce the results from the paper, use the following hyperparameters settings:

Hyperparameters Settings

Dataset batch_size edim rdim k n_i TR_ranks dr lr input_d hidden_d
WikiPeople-3 128 50 50 4 50 50 0.995 0.0009267003174594345 0.3740776415163665 0.45137914784181227
WikiPeople-4 128 25 25 5 25 40 0.995 0.006701566797680926 0.46694419227220374 0.18148844341064124
JF17K-3 128 50 50 4 50 50 0.99 0.0008658318809880197 0.12747824547053027 0.501929359180091
JF17K-4 128 25 25 5 25 40 0.995 0.0006071265071591076 0.010309222253012645 0.43198147413900445

Link Prediction Results on binary Relational KBs (KGs)

Dataset MRR Hits@10 Hits@3 Hits@1
WN18 0.948 0.954 0.950 0.945
FB15k 0.824 0.888 0.847 0.787

Running a model

CUDA_VISIBLE_DEVICES=0 python main.py --dataset dataset --num_iterations 200 --batch_size batch_size 
--edim edim --rdim rdim --k k --n_i n_i --TR_ranks TR_ranks --dr dr --lr lr --input_dropout --input_d 
--hidden_dropout1 hidden_d1 --hidden_dropout2 hidden_d2

Available datasets are:

WN18, FB15K

Hyperparameters Settings

Dataset batch_size edim rdim k n_i TR_ranks dr lr input_d hidden_d1 hidden_d2
WN18 128 200 200 3 200 50 0.995 0.0005 0.2 0.1 0.2
FB15k 512 200 200 3 200 50 0.995 0.0005 0.3 0.5 0.0

Requirements

python     3.6.8
numpy     1.16.3.1
pytorch    1.0.1

Reference

@inproceddings{liu2020getd,
	title 	  = {Generalizing Tensor Decomposition for N-ary Relational Knowledge Base},
	author	  = {Liu, Yu and Yao, Quanming and Li, Yong},
	booktitle = {The World Wide Web Conference},
	year      = {2020},
}

Acknowledgement

The codes of this paper are based on the codes of TuckER (https://github.com/ibalazevic/TuckER) . We appreciate TuckER's codes and thank the authors of TuckER.