Knowledge Base Completion (kbc)

This code reproduces results in Canonical Tensor Decomposition for Knowledge Base Completion (ICML 2018).

Installation

Create a conda environment with pytorch cython and scikit-learn :

conda create --name kbc_env python=3.7
source activate kbc_env
conda install --file requirements.txt -c pytorch

Then install the kbc package to this environment

python setup.py install

Datasets

To download the datasets, go to the kbc/scripts folder and run:

chmod +x download_data.sh
./download_data.sh

Once the datasets are download, add them to the package data folder by running :

python kbc/process_datasets.py

This will create the files required to compute the filtered metrics.

Running the code

Reproduce the results below with the following command :

python kbc/learn.py --dataset FB15K --model ComplEx --rank 500 --optimizer
Adagrad --learning_rate 1e-1 --batch_size 1000 --regularizer N3 --reg 1e-2
 --max_epochs 100 --valid 5

Results

In addition to the results in the paper, here are the performances of ComplEx regularized with the weighted N3 on several datasets, for several dimensions. We use an init scale of 1e-3, a learning rate of 0.1, a batch size of 1000 and 100 max epochs unless specified otherwise. We use the Adagrad optimizer.

FB15k

For rank 2000 : learning rate 1e-2, batch-size 100, max epochs 200.

rank	5	25	50	100	500	2000
MRR	0.36	0.61	0.78	0.83	0.84	0.86
H@1	0.27	0.52	0.73	0.79	0.80	0.83
H@3	0.41	0.67	0.81	0.85	0.87	0.87
H@10	0.55	0.77	0.86	0.89	0.91	0.91


reg	1e-5	1e-5	1e-5	7.5e-4	1e-2	2.5e-3
#Params	163k	815k	1.630M	3.259M	1.630M	65.184M

WN18

Max Epochs : 20

rank	5	8	16	25	50	100	500	2000
MRR	0.19	0.45	0.92	0.94	0.95	0.95	0.95	0.95
H@1	0.14	0.37	0.91	0.94	0.94	0.94	0.94	0.94
H@3	0.20	0.50	0.93	0.94	0.95	0.95	0.95	0.95
H@10	0.29	0.60	0.94	0.95	0.95	0.95	0.96	0.96


reg	1e-3	5e-4	5e-4	1e-3	5e-3	5e-2	5e-2	5e-2
#Params	410k	656k	1.311M	2.049M	4.098M	8.196M	40.979M	163.916M

FB15K-237

Batch Size : 100 (1000 for rank 1000)

rank	5	25	50	100	500	1000	2000
MRR	0.28	0.33	0.34	0.35	0.36	0.37	0.37
H@1	0.20	0.24	0.25	0.26	0.27	0.27	0.27
H@3	0.31	0.36	0.37	0.39	0.40	0.40	0.40
H@10	0.44	0.51	0.52	0.54	0.56	0.56	0.56


reg	5e-4	5e-2	5e-2	5e-2	5e-2	5e-2	5e-2
#Params	150k	751k	1.502M	3.003M	15.015M	30.030M	60.060M

WN18RR

Batch Size : 100 (1000 for rank 8)

rank	5	8	16	25	50	100	500	2000
MRR	0.26	0.36	0.42	0.44	0.46	0.47	0.49	0.49
H@1	0.20	0.38	0.39	0.41	0.43	0.43	0.44	0.44
H@3	0.29	0.38	0.42	0.45	0.47	0.49	0.50	0.50
H@10	0.36	0.41	0.46	0.49	0.52	0.56	0.58	0.58


reg	5e-4	5e-4	5e-2	1e-1	1e-1	1e-1	1e-1	1e-1
#Params	410k	655k	1.311M	2.048M	4.097M	8.193M	40.975M	163.860M

YAGO3-10

rank	5	16	25	50	100	500	1000
MRR	0.15	0.34	0.46	0.54	0.56	0.57	0.58
H@1	0.10	0.26	0.38	0.47	0.49	0.50	0.50
H@3	0.16	0.37	0.50	0.58	0.60	0.62	0.62
H@10	0.25	0.50	0.60	0.67	0.69	0.71	0.71


reg	1e-3	1e-4	5e-3	5e-3	5e-3	5e-3	5e-3
#Params	1.233M	3.944M	6.163M	12.326M	24.652M	123.262M	246.524M

License

kbc is CC-BY-NC licensed, as found in the LICENSE file.

facebookresearch / kbc

readme