KGEval
A framework for evaluating Knowledge Graph Embedding Models in a fine-grained manner.
The framework and experimental results are described in Ben Rim et al. 2021 (Outstanding Paper Award, AKBC 2021).
Instructions
Create a virtual environment
virtualenv -p python3.6 eval_env
source eval_env/bin/activate
pip install -r requirements.txt
Download data
In the main folder, run:
source data/download.sh
Download model
If you want to test the framework immediately, you can download pre-trained Pykeen models by running:
source download_models.sh
Generate behavioral tests
Symmetry Tests
Can choose --dataset FB15K237, WN18RR, YAGO310
python tests/run.py --dataset FB15K237 --mode generate --capability symmetry
This should result into the following output, and the files for each test set will be added under behavioral_tests\dataset\symmetry:
2021-10-03 23:37:35,060 - [INFO] - Preparing test sets for the dataset FB15K237
2021-10-03 23:37:37,621 - [INFO] - ########################## <----TRAIN---> ############################
2021-10-03 23:37:37,621 - [INFO] - 0 repetitions removed
2021-10-03 23:37:37,621 - [INFO] - 272115 triples remaining in train set
2021-10-03 23:37:37,621 - [INFO] - 6778 symmetric triples found in train set
2021-10-03 23:37:37,786 - [INFO] - ########################## <----TEST---> ############################
2021-10-03 23:37:37,786 - [INFO] - 0 repetitions removed
2021-10-03 23:37:37,786 - [INFO] - 20466 triples remaining in test set
2021-10-03 23:37:37,786 - [INFO] - 113 symmetric triples found in test set
2021-10-03 23:37:37,806 - [INFO] - ########################## <----VALID---> ############################
2021-10-03 23:37:37,806 - [INFO] - 0 repetitions removed
2021-10-03 23:37:37,806 - [INFO] - 17535 triples remaining in valid set
2021-10-03 23:37:37,806 - [INFO] - 113 symmetric triples found in valid set
2021-10-03 23:37:39,106 - [INFO] - #################### <---TEST SET 1: MEMORIZATION ---> ##########################
2021-10-03 23:37:39,106 - [INFO] - There are 5470 entries in the memorization set (occur in both directions)
2021-10-03 23:37:39,106 - [INFO] - #################### <---TEST SET 2: ONE DIRECTION SEEN ---> ##########################
2021-10-03 23:37:39,106 - [INFO] - There are 1308 entries not shown in both directions (to be reversed for testing)
2021-10-03 23:37:39,836 - [INFO] - #################### <--- SYMMETRIC RELATIONS ---> ##########################
2021-10-03 23:37:39,836 - [INFO] - TRAIN SET contains 6778 symmetric entries
2021-10-03 23:37:39,836 - [INFO] - TEST SET contains 113 symmetric entries with 113 not in training
2021-10-03 23:37:39,836 - [INFO] - VALID SET contains 113 symmetric entries with 113 not in training
2021-10-03 23:37:39,839 - [INFO] - #################### <---TEST SET 3: UNSEEN INSTANCES ---> ##########################
2021-10-03 23:37:39,840 - [INFO] - There are 226 entries that are not seen in any direction in training
2021-10-03 23:37:40,267 - [INFO] - #################### <---TEST SET 4: ASYMMETRY ---> ##########################
2021-10-03 23:37:40,267 - [INFO] - There are 3000 asymmetric entries in test set added to test 4
Hierarchy Tests
Only available for FB15K237 dataset
python tests/run.py --dataset FB15K237 --mode generate --capability hierarchy
The output should be and will be available under behavioral_tests/dataset/hierarchy/, the naming of the files corresponds to triples where the tail belongs to a specified level. For example, 1.txt contains triples where the tail has a type of level 1 in the entity type hierarchy :
2021-10-04 01:38:13,517 - [INFO] - Results of Hierarchy Behavioral Tests for FB15K237
2021-10-04 01:38:20,367 - [INFO] - <--------------- Entity Hiararchy statistics ----------------->
2021-10-04 01:38:20,568 - [INFO] - Level 0 contains 1 types and 3415 triples
2021-10-04 01:38:20,887 - [INFO] - Level 1 contains 66 types and 2006 triples
2021-10-04 01:38:20,900 - [INFO] - Level 2 contains 136 types and 4273 triples
2021-10-04 01:38:20,913 - [INFO] - Level 3 contains 213 types and 3560 triples
2021-10-04 01:38:20,923 - [INFO] - Level 4 contains 262 types and 3369 triples
Run Tests (pykeen models)
Symmetry behavioral tests on distmult or rotate:
python tests/run.py --dataset FB15K237 --mode test --model_name rotate
The output will be printed as shown below, and will also be available in the results folder under dataset/symmetry:
2021-10-04 14:00:57,100 - [INFO] - Starting test1 with rotate model
2021-10-04 14:03:23,249 - [INFO] - On test1, MR: 1.2407678244972578, MRR: 0.9400152688974949, [email protected]: 0.9014624953269958, [email protected]: 0.988482654094696, [email protected]: 0.9965264797210693
2021-10-04 14:03:23,249 - [INFO] - Starting test2 with rotate model
2021-10-04 14:04:15,614 - [INFO] - On test2, MR: 23.446483180428135, MRR: 0.4409348919640765, [email protected]: 0.30351680517196655, [email protected]: 0.5894495248794556, [email protected]: 0.7025994062423706
2021-10-04 14:04:15,614 - [INFO] - Starting test3 with rotate model
2021-10-04 14:04:25,364 - [INFO] - On test3, MR: 1018.9469026548672, MRR: 0.04786047740344238, [email protected]: 0.008849557489156723, [email protected]: 0.06194690242409706, [email protected]: 0.12389380484819412
2021-10-04 14:04:25,365 - [INFO] - Starting test4 with rotate model
2021-10-04 14:05:38,900 - [INFO] - On test4, MR: 4901.459, MRR: 0.07606098649786266, [email protected]: 0.9496666789054871, [email protected]: 0.893666684627533, [email protected]: 0.8823333382606506
Hierarchy behavioral tests on distmult or rotate:
python tests/run.py --dataset FB15K237 --mode test --capability hierarchy --model_name rotate
Run Tests on other models and other frameworks
(To be added)
3 Aug 25, 2022
13 Oct 24, 2022
1 Jul 08, 2022
41 Dec 08, 2022
8 Jan 06, 2023
1 Jan 13, 2022
1.8k Dec 21, 2022
11 Oct 31, 2022
2 Jan 09, 2022
109 Jan 04, 2023
18 Jan 11, 2022
186 Dec 29, 2022
29 Nov 26, 2022
1 Oct 18, 2022
29 Jan 05, 2023
1 Dec 03, 2021
37 Dec 07, 2022
7 Dec 25, 2022
24 Dec 10, 2022
10 Dec 27, 2022