Skip to content

Latest commit

 

History

History
389 lines (331 loc) · 22.5 KB

README.md

File metadata and controls

389 lines (331 loc) · 22.5 KB
Raw

Downloads Downloads Coverage Pypi Docs Python  

Ontolearn

Ontolearn: Learning OWL Class Expressions

Ontolearn is an open-source software library for learning owl class expressions at large scale.

Given positive and negative OWL named individual examples $E^+$ and $E^-$, learning OWL Class expression problem refers to the following supervised Machine Learning problem

$$\forall p \in E^+\ \mathcal{K} \models H(p) \wedge \forall n \in E^-\ \mathcal{K} \not \models H(n).$$

To tackle this supervised learning problem, ontolearn offers many symbolic, neuro-symbolic and deep learning based Learning algorithms:

Find more in the Documentation.

Installation

pip install ontolearn

or

git clone https://github.com/dice-group/Ontolearn.git 
# To create a virtual python env with conda 
conda create -n venv python=3.10.14 --no-default-packages && conda activate venv && pip install -e .
# To download knowledge graphs
wget https://files.dice-research.org/projects/Ontolearn/KGs.zip -O ./KGs.zip && unzip KGs.zip
# To download learning problems
wget https://files.dice-research.org/projects/Ontolearn/LPs.zip -O ./LPs.zip && unzip LPs.zip

Learning OWL Class Expressions

from ontolearn.learners import TDL
from ontolearn.triple_store import TripleStore
from ontolearn.knowledge_base import KnowledgeBase
from ontolearn.learning_problem import PosNegLPStandard
from owlapy.owl_individual import OWLNamedIndividual
from owlapy import owl_expression_to_sparql, owl_expression_to_dl
# (1) Initialize Triplestore or KnowledgeBase
# sudo docker run -p 3030:3030 -e ADMIN_PASSWORD=pw123 stain/jena-fuseki
# Login http://localhost:3030/#/ with admin and pw123 and upload KGs/Family/family.owl
# kb = TripleStore(url="http://localhost:3030/family")
kb = KnowledgeBase(path="KGs/Family/father.owl")
# (2) Initialize a learner.
model = TDL(knowledge_base=kb, use_nominals=True)
# (3) Define a description logic concept learning problem.
lp = PosNegLPStandard(pos={OWLNamedIndividual("http://example.com/father#stefan")},
                      neg={OWLNamedIndividual("http://example.com/father#heinz"),
                           OWLNamedIndividual("http://example.com/father#anna"),
                           OWLNamedIndividual("http://example.com/father#michelle")})
# (4) Learn description logic concepts best fitting (3).
h = model.fit(learning_problem=lp).best_hypotheses()
print(h) 
print(owl_expression_to_dl(h))
print(owl_expression_to_sparql(expression=h)) 
"""
OWLObjectSomeValuesFrom(property=OWLObjectProperty(IRI('http://example.com/father#','hasChild')),filler=OWLObjectOneOf((OWLNamedIndividual(IRI('http://example.com/father#','markus')),)))

∃ hasChild.{markus}

SELECT
 DISTINCT ?x WHERE { 
?x <http://example.com/father#hasChild> ?s_1 . 
 FILTER ( ?s_1 IN ( 
<http://example.com/father#markus>
 ) )
 }
"""
print(model.classification_report)
"""
Classification Report: Negatives: -1 and Positives 1 
              precision    recall  f1-score   support

    Negative       1.00      1.00      1.00         3
    Positive       1.00      1.00      1.00         1

    accuracy                           1.00         4
   macro avg       1.00      1.00      1.00         4
weighted avg       1.00      1.00      1.00         4
"""

Learning OWL Class Expressions over DBpedia

from ontolearn.learners import TDL, Drill
from ontolearn.triple_store import TripleStore
from ontolearn.learning_problem import PosNegLPStandard
from owlapy.owl_individual import OWLNamedIndividual
from owlapy import owl_expression_to_sparql, owl_expression_to_dl
from ontolearn.utils.static_funcs import save_owl_class_expressions
# (1) Initialize Triplestore
kb = TripleStore(url="https://dbpedia.data.dice-research.org/sparql")
# (3) Initialize a learner.
model = Drill(knowledge_base=kb) #  or  TDL(knowledge_base=kb)
# (4) Define a description logic concept learning problem.
lp = PosNegLPStandard(pos={OWLNamedIndividual("http://dbpedia.org/resource/Angela_Merkel")},
                      neg={OWLNamedIndividual("http://dbpedia.org/resource/Barack_Obama")})
# (5) Learn description logic concepts best fitting (4).
h = model.fit(learning_problem=lp).best_hypotheses()
print(h)
print(owl_expression_to_dl(h))
print(owl_expression_to_sparql(expression=h))
save_owl_class_expressions(expressions=h,path="#owl_prediction")

Fore more please refer to the examples folder.

ontolearn-webservice

Click me!

Load an RDF knowledge graph

ontolearn-webservice --path_knowledge_base KGs/Mutagenesis/mutagenesis.owl

or launch a Tentris instance https://github.com/dice-group/tentris over Mutagenesis.

ontolearn-webservice --endpoint_triple_store http://0.0.0.0:9080/sparql

The below code trains DRILL with 6 randomly generated learning problems provided that path_to_pretrained_drill does not lead to a directory containing pretrained DRILL. Thereafter, trained DRILL is saved in the directory path_to_pretrained_drill. Finally, trained DRILL will learn an OWL class expression.

import json
import requests
with open(f"LPs/Mutagenesis/lps.json") as json_file:
    learning_problems = json.load(json_file)["problems"]
for str_target_concept, examples in learning_problems.items():
    response = requests.get('http://0.0.0.0:8000/cel',
                            headers={'accept': 'application/json', 'Content-Type': 'application/json'},
                            json={"pos": examples['positive_examples'],
                                  "neg": examples['negative_examples'],
                                  "model": "Drill",
                                  "path_embeddings": "mutagenesis_embeddings/Keci_entity_embeddings.csv",
                                  "path_to_pretrained_drill": "pretrained_drill",
                                  # if pretrained_drill exists, upload, otherwise train one and save it there
                                  "num_of_training_learning_problems": 2,
                                  "num_of_target_concepts": 3,
                                  "max_runtime": 60000,  # seconds
                                  "iter_bound": 1  # number of iterations/applied refinement opt.
                                  })
    print(response.json())  # {'Prediction': '∀ hasAtom.(¬Nitrogen-34)', 'F1': 0.7283582089552239, 'saved_prediction': 'Predictions.owl'}

TDL (a more scalable learner) can also be used as follows

import json
import requests
response = requests.get('http://0.0.0.0:8000/cel',
                        headers={'accept': 'application/json', 'Content-Type': 'application/json'},
                        json={"pos": examples['positive_examples'],
                              "neg": examples['negative_examples'],
                              "model": "TDL"})
print(response.json())

NCES (another scalable learner). The following will first train NCES if the provided path path_to_pretrained_nces does not exist

import json
import requests
with open(f"LPs/Mutagenesis/lps.json") as json_file:
    learning_problems = json.load(json_file)["problems"]
## This trains NCES before solving the provided learning problems. Expect poor performance for this number of epochs, and this training data size.
## If GPU is available, set `num_of_training_learning_problems` t0 10_000 or more. Set `nces_train_epochs` to 300 or more, and increase `nces_batch_size`.
for str_target_concept, examples in learning_problems.items():
    response = requests.get('http://0.0.0.0:8000/cel',
                            headers={'accept': 'application/json', 'Content-Type': 'application/json'},
                            json={"pos": examples['positive_examples'],
                                  "neg": examples['negative_examples'],
                                  "model": "NCES",
                                  "path_embeddings": "mutagenesis_embeddings/Keci_entity_embeddings.csv",
                                  "path_to_pretrained_nces": None,
                                  # if pretrained_nces exists, load weghts, otherwise train one and save it
                                  "num_of_training_learning_problems": 100,
                                  "nces_train_epochs": 5,
                                  "nces_batch_size": 16
                                  })
    print(response.json())

Now this will use pretrained weights for NCES

import json
import requests
with open(f"LPs/Mutagenesis/lps.json") as json_file:
    learning_problems = json.load(json_file)["problems"]
for str_target_concept, examples in learning_problems.items():
    response = requests.get('http://0.0.0.0:8000/cel',
                            headers={'accept': 'application/json', 'Content-Type': 'application/json'},
                            json={"pos": examples['positive_examples'],
                                  "neg": examples['negative_examples'],
                                  "model": "NCES",
                                  "path_embeddings": "./NCESData/mutagenesis/embeddings/ConEx_entity_embeddings.csv",
                                  "path_to_pretrained_nces": "./NCESData/mutagenesis/trained_models/",
                                  # if pretrained_nces exists, load weghts, otherwise train one and save it
                                  "num_of_training_learning_problems": 100,
                                  "nces_train_epochs": 5,
                                  "nces_batch_size": 16
                                  })
    print(response.json())

Benchmark Results

To see the results 
# To download learning problems. # Benchmark learners on the Family benchmark dataset with benchmark learning problems.
wget https://files.dice-research.org/projects/Ontolearn/LPs.zip -O ./LPs.zip && unzip LPs.zip

10-Fold Cross Validation Family Benchmark Results

Here we apply 10-fold cross validation technique on each benchmark learning problem with max runtime of 60 seconds to measure the training and testing performance of learners. In the evaluation, from a given single learning problem (a set of positive and negative examples), a learner learns an OWL Class Expression (H) on a given 9 fold of positive and negative examples. To compute the training performance, We compute F1-score of H train positive and negative examples. To compute the test performance, we compute F1-score of H w.r.t. test positive and negative examples.

# To download learning problems and benchmark learners on the Family benchmark dataset with benchmark learning problems.
python examples/concept_learning_cv_evaluation.py --kb ./KGs/Family/family-benchmark_rich_background.owl --lps ./LPs/Family/lps_difficult.json --path_of_nces_embeddings ./NCESData/family/embeddings/ConEx_entity_embeddings.csv --path_of_clip_embeddings ./CLIPData/family/embeddings/ConEx_entity_embeddings.csv --max_runtime 60 --report family_results.csv

In the following python script, the results are summarized and the markdown displayed below generated.

import pandas as pd
df=pd.read_csv("family_results.csv").groupby("LP").mean()
print(df[[col for col in df if col.startswith('Test-F1') or col.startswith('RT')]].to_markdown(floatfmt=".3f"))

Note that DRILL is untrained and we simply used accuracy driven heuristics to learn an OWL class expression.

Below, we report the average test F1 score and the average runtimes of learners.

LP Test-F1-OCEL RT-OCEL Test-F1-CELOE RT-CELOE Test-F1-Evo RT-Evo Test-F1-DRILL RT-DRILL Test-F1-TDL RT-TDL Test-F1-NCES RT-NCES Test-F1-CLIP RT-CLIP
Aunt 0.614 13.697 0.855 13.697 0.978 5.278 0.811 60.351 0.956 0.118 0.812 1.168 0.855 14.059
Cousin 0.712 10.846 0.789 10.846 0.993 3.311 0.701 60.485 0.820 0.176 0.677 1.050 0.779 9.050
Grandgranddaughter 1.000 0.013 1.000 0.013 1.000 0.426 0.980 17.486 1.000 0.050 1.000 0.843 1.000 0.639
Grandgrandfather 1.000 0.897 1.000 0.897 1.000 0.404 0.947 55.728 0.947 0.059 0.927 0.902 1.000 0.746
Grandgrandmother 1.000 4.173 1.000 4.173 1.000 0.442 0.893 50.329 0.947 0.060 0.927 0.908 1.000 0.817
Grandgrandson 1.000 1.632 1.000 1.632 1.000 0.452 0.931 60.358 0.911 0.070 0.911 1.050 1.000 0.939
Uncle 0.876 16.244 0.891 16.244 0.964 4.516 0.876 60.416 0.933 0.098 0.891 1.256 0.928 17.682
LP Train-F1-OCEL Train-F1-CELOE Train-F1-Evo Train-F1-DRILL Train-F1-TDL Train-F1-NCES Train-F1-CLIP
Aunt 0.835 0.918 0.995 0.837 1.000 0.804 0.918
Cousin 0.746 0.796 1.000 0.732 1.000 0.681 0.798
Grandgranddaughter 1.000 1.000 1.000 1.000 1.000 1.000 1.000
Grandgrandfather 1.000 1.000 1.000 0.968 1.000 0.973 1.000
Grandgrandmother 1.000 1.000 1.000 0.975 1.000 0.939 1.000
Grandgrandson 1.000 1.000 1.000 0.962 1.000 0.927 1.000
Uncle 0.904 0.907 0.996 0.908 1.000 0.884 0.940

10-Fold Cross Validation Mutagenesis Benchmark Results

python examples/concept_learning_cv_evaluation.py --kb ./KGs/Mutagenesis/mutagenesis.owl --lps ./LPs/Mutagenesis/lps.json --path_of_nces_embeddings ./NCESData/mutagenesis/embeddings/ConEx_entity_embeddings.csv --path_of_clip_embeddings ./CLIPData/mutagenesis/embeddings/ConEx_entity_embeddings.csv --max_runtime 60 --report mutagenesis_results.csv
LP Train-F1-OCEL Test-F1-OCEL RT-OCEL Train-F1-CELOE Test-F1-CELOE RT-CELOE Train-F1-Evo Test-F1-Evo RT-Evo Train-F1-DRILL Test-F1-DRILL RT-DRILL Train-F1-TDL Test-F1-TDL RT-TDL Train-F1-NCES Test-F1-NCES RT-NCES Train-F1-CLIP Test-F1-CLIP RT-CLIP
NotKnown 0.916 0.918 60.705 0.916 0.918 60.705 0.975 0.970 51.870 0.809 0.804 60.140 1.000 0.852 13.569 0.717 0.718 3.784 0.916 0.918 26.312

10-Fold Cross Validation Carcinogenesis Benchmark Results

python examples/concept_learning_cv_evaluation.py --kb ./KGs/Carcinogenesis/carcinogenesis.owl --lps ./LPs/Carcinogenesis/lps.json --path_of_nces_embeddings ./NCESData/carcinogenesis/embeddings/ConEx_entity_embeddings.csv --path_of_clip_embeddings ./CLIPData/carcinogenesis/embeddings/ConEx_entity_embeddings.csv --max_runtime 60 --report carcinogenesis_results.csv
LP Train-F1-OCEL Test-F1-OCEL RT-OCEL Train-F1-CELOE Test-F1-CELOE RT-CELOE Train-F1-Evo Test-F1-Evo RT-Evo Train-F1-DRILL Test-F1-DRILL RT-DRILL Train-F1-TDL Test-F1-TDL RT-TDL Train-F1-NCES Test-F1-NCES RT-NCES Train-F1-CLIP Test-F1-CLIP RT-CLIP
NOTKNOWN 0.737 0.711 62.048 0.740 0.701 62.048 0.822 0.628 64.508 0.740 0.707 60.120 1.000 0.616 5.196 0.705 0.704 4.157 0.740 0.701 48.475

Benchmark Results on DBpedia. Results are based on the training examples only

python examples/owl_class_expression_learning_dbpedia.py --model Drill && python examples/owl_class_expression_learning_dbpedia.py --model TDL
LP-Type Train-F1-Drill RT-Drill Train-F1-TDL RT-TDL
OWLObjectAllValuesFrom 0.438 240.331 1.000 206.288
OWLObjectIntersectionOf 0.213 202.558 0.717 91.660
OWLObjectUnionOf 0.546 187.144 0.967 129.700

Development

To see the results

Creating a feature branch refactoring from development branch

git branch refactoring develop

Each feature branch must be merged to develop branch. To this end, the tests must run without a problem:

# To download knowledge graphs
wget https://files.dice-research.org/projects/Ontolearn/KGs.zip -O ./KGs.zip && unzip KGs.zip
# To download learning problems
wget https://files.dice-research.org/projects/Ontolearn/LPs.zip -O ./LPs.zip && unzip LPs.zip
# Download weights for some model for few tests
wget https://files.dice-research.org/projects/NCES/NCES_Ontolearn_Data/NCESData.zip -O ./NCESData.zip && unzip NCESData.zip && rm NCESData.zip
wget https://files.dice-research.org/projects/Ontolearn/CLIP/CLIPData.zip && unzip CLIPData.zip && rm CLIPData.zip 
pytest -p no:warnings -x # Running 76 tests takes ~ 17 mins

References

Currently, we are working on our manuscript describing our framework. If you find our work useful in your research, please consider citing the respective paper:

# DRILL
@inproceedings{demir2023drill,
  author = {Demir, Caglar and Ngomo, Axel-Cyrille Ngonga},
  booktitle = {The 32nd International Joint Conference on Artificial Intelligence, IJCAI 2023},
  title = {Neuro-Symbolic Class Expression Learning},
  url = {https://www.ijcai.org/proceedings/2023/0403.pdf},
 year={2023}
}

# NCES2
@inproceedings{kouagou2023nces2,
author={Kouagou, N'Dah Jean and Heindorf, Stefan and Demir, Caglar and Ngonga Ngomo, Axel-Cyrille},
title={Neural Class Expression Synthesis in ALCHIQ(D)},
url = {https://papers.dice-research.org/2023/ECML_NCES2/NCES2_public.pdf},
booktitle={Machine Learning and Knowledge Discovery in Databases},
year={2023},
publisher={Springer Nature Switzerland},
address="Cham"
}

# NCES
@inproceedings{kouagou2023neural,
  title={Neural class expression synthesis},
  author={Kouagou, N’Dah Jean and Heindorf, Stefan and Demir, Caglar and Ngonga Ngomo, Axel-Cyrille},
  booktitle={European Semantic Web Conference},
  pages={209--226},
  year={2023},
  publisher={Springer Nature Switzerland}
}

# EvoLearner
@inproceedings{heindorf2022evolearner,
  title={Evolearner: Learning description logics with evolutionary algorithms},
  author={Heindorf, Stefan and Bl{\"u}baum, Lukas and D{\"u}sterhus, Nick and Werner, Till and Golani, Varun Nandkumar and Demir, Caglar and Ngonga Ngomo, Axel-Cyrille},
  booktitle={Proceedings of the ACM Web Conference 2022},
  pages={818--828},
  year={2022}
}


# CLIP
@inproceedings{kouagou2022learning,
  title={Learning Concept Lengths Accelerates Concept Learning in ALC},
  author={Kouagou, N’Dah Jean and Heindorf, Stefan and Demir, Caglar and Ngonga Ngomo, Axel-Cyrille},
  booktitle={European Semantic Web Conference},
  pages={236--252},
  year={2022},
  publisher={Springer Nature Switzerland}
}

In case you have any question, please contact: caglar.demir@upb.de or caglardemir8@gmail.com