EnhancedRuleLearning

We implement an approach for improving rule learning via embedding-based KG enrichement

Installation

Clone this repository:

https://github.com/Jean-KOUAGOU/EnhancedRuleLearning.git

First install Anaconda3, then all required librairies by running the following:

conda env create -f environment.yml

A conda environment (enhancerl) will be created. Next activate the environment: conda activate enhancerl

An error may occur during installation of pykeen. In such a case, we recommend using pip install pykeen in the newly created environment "enhancerl".

Install Java (version 8 or higher) in the environment (enhancerl): conda install -c bioconda java-jdk

Download datasets and pretrained models by running ./download_data

Reproduce the reported results

1. Statistics of the generated rules using our approach: Make sure a jupyter kernel with the required libraries in enhancerl is installed. Open the notebook (analyzing mined rules)[analyzing_mined_rules.ipynb], activate the previously kernel and run all cells

2. Performance of the generated rules on link prediction on test sets: Open a terminal and run python evaluate-rules-hits.py with the following parameter options:

  --kgs {wn18rr,fb15k,carcinogenesis,mutagenesis,openbiolink,yago3,drkg} [{wn18rr,fb15k,carcinogenesis,mutagenesis,openbiolink,yago3,drkg} ...]
                        Knowledge graph names, defaulting to wn18rr
  --system {amie,AMIE,anyburl,AnyBURL}
                        System to execute, defaulting to amie
  --models {TransE,Distmult,Rotate,transe,distmult,rotate} [{TransE,Distmult,Rotate,transe,distmult,rotate} ...]
                        Embedding model(s), defaulting to None

Running rule mining systems

1. Learning rules: Open a terminal and run python execute-system.py. Parameter setting is the same as above

2. Configuration files: Check whether configuration files are already present under data/{System-Folder}. Otherwise, run python create-config.py, where {System-Folder} is one of AMIE, DistAMIE, RotAMIE, AnyBURL

3. Applying mined rules: Open a terminal and run python apply-rules.py. Parameter setting is also the same as above (i.e., same as in Reproduce the reported results)

Training embeddings

Open a terminal in EnhancedRuleLearning and run python src/embed_kgs_pykeen.py. Use python src/embed_kgs_pykeen.py -h to view all available options

Inferring new triples for KG completion

Open a terminal in the main repository

• Predict new links: python src/predict_new_links.py with the following options:

  --models {transe,distmult,rotate} [{transe,distmult,rotate} ...]
                        The embedding model used, e.g. TransE
  --kgs KGS [KGS ...]   The knowledge graph name
  --target_relations TARGET_RELATIONS [TARGET_RELATIONS ...]
                        Relation names to use for link prediction
  --num_entities NUM_ENTITIES
                        The number of entities to select
  --random_seed RANDOM_SEED
                        Random seed for random entity selection for link prediction
  --num_relations NUM_RELATIONS
                        The maximum number of relations to use in case none was given
  --topk_triples TOPK_TRIPLES
                        Top 50 highest scoring triples
  --store_triples [STORE_TRIPLES]
                        Whether to store the newly predicted triples

Results

Runtime

Link prediction performance using the mined rules

	Hits@1	Hits@3	MRR
	AMIE+
WN18RR	0.390	0.395	0.398
CARCINOGENESIS	0.216	0.245	0.248
MUTAGENESIS	0.217	0.244	0.244
FB15K-237	0.180	0.284	0.389
YAGO3-10	0.244	0.333	0.408
DRKG	0.147	0.172	0.215
OPENBIOLINK	0.038	0.081	0.158
	Ours (with RotatE)
WN18RR	0.384	0.390	0.392
CARCINOGENESIS	0.220	0.248	0.252
MUTAGENESIS	0.182	0.215	0.227
FB15K-237	0.177	0.280	0.385
YAGO3-10	0.243	0.331	0.405
DRKG	0.147	0.172	0.215
OPENBIOLINK	0.059	0.129	0.258
	AnyBURL
WN18RR	0.469	0.527	0.594
CARCINOGENESIS	0.483	0.533	0.630
MUTAGENESIS	0.477	0.547	0.639
FB15K-237	0.278	0.424	0.569
YAGO3-10	0.372	0.457	0.531
DRKG	0.064	0.075	0.090
OPENBIOLINK	0.009	0.016	0.026