Concept Learning

This is a guide to show how to use a concept learner to generate hypotheses for a target concept in an ontology. In this guide we will show how to use the following concept learners of Ontolearn library:

The other concept learners are not covered here in details, but we have provided examples for them. Check the jupyter notebook files as well as other example scripts for the corresponding learner inside the examples folder (direct links are given at the end of this guide).

It is worth mentioning that NCES2 and NERO are not yet implemented in Ontolearn, but they will be soon.

Expressiveness

Evolearner → ALCQ(D).

DRILL → ALC

NCES → ALC

NCES2 → ALCHIQ(D)

NERO → ALC

CLIP → ALC

CELOE and OCEL → ALC


The three algorithms that we mentioned in the beginning are similar in execution, for that reason, we are describing them in a general manner. To test them separately see Quick try-out. Each algorithm may have different available configuration. However, at minimum, they require a knowledge base and a learning problem.

Let’s see the prerequisites needed to run the concept learners:

Prerequisites

Before configuring and running an algorithm, we recommend you to store the dataset path that ends with .owl and the IRIs as string of the learning problem instances in a json file as shown below. The learning problem is further divided in positive and negative examples. We have saved ourselves some hardcoded lines which we can now simply access by loading the json file. Below is an example file that we are naming synthetic_problems.json showing how should it look:

{  
  "data_path": "../KGs/Family/family-benchmark_rich_background.owl",  
  "learning_problem": {
    "positive_examples": [  
    "http://www.benchmark.org/family#F2F28",  
    "http://www.benchmark.org/family#F2F36",  
    "http://www.benchmark.org/family#F3F52"  
    ],  
    "negative_examples": [  
    "http://www.benchmark.org/family#F6M69",  
    "http://www.benchmark.org/family#F6M100",  
    "http://www.benchmark.org/family#F2F30"  
    ]
  }  
}

We are considering that you are trying this script inside examples folder, and therefore we have stored the ontology path like that.

Note: The KGs directory contains datasets, and it’s not part of the project. They have to be downloaded first, see Download External Files. You can also download some ready to use learning problem json files by clicking here.

Configuring Input Parameters

Before starting with the configuration you can enable logging to see the logs which give insights about the main processes of the algorithm:

from ontolearn.utils import setup_logging

setup_logging()

We then start by loading the synthetic_problems.json where we have stored the knowledge base path and the learning problems in the variable settings:

import json

with open('synthetic_problems.json') as json_file:    
    settings = json.load(json_file)

Load the ontology

Load the ontology by simply creating an instance of the class KnowledgeBase and passing the ontology path stored under data_path property of settings:

from ontolearn.knowledge_base import KnowledgeBase

kb = KnowledgeBase(path=settings['data_path'])

Configure the Learning Problem

The Structured Machine Learning implemented in our Ontolearn library is working with a type of supervised learning. One of the first things to do after loading the Ontology to a KnowledgeBase object is thus to define the learning problem for which the learning algorithm is trying to generate hypothesis (class expressions).

First and foremost, load the learning problem examples from the json file into sets as shown below:

positive_examples = set(settings['learning_problem']['positive_examples'])  
negative_examples = set(settings['learning_problem']['negative_examples'])

In Ontolearn you represent the learning problem as an object of the class PosNegLPStandard which has two parameters pos and neg respectively for the positive and negative examples. These parameters are of type set[OWLNamedIndividual]. We create these sets by mapping each individual (stored as string) from the set positive_examples and negative_examples to OWLNamedIndividual:

from ontolearn.learning_problem import PosNegLPStandard
from owlapy.owl_individual import IRI, OWLNamedIndividual

typed_pos = set(map(OWLNamedIndividual, map(IRI.create, positive_examples)))
typed_neg = set(map(OWLNamedIndividual, map(IRI.create, negative_examples)))
lp = PosNegLPStandard(pos=typed_pos, neg=typed_neg)

To construct an OWLNamedIndividual object an IRI is required as an input. You can simply create an IRI object by calling the static method create and passing the IRI as a string.

Configuring & Executing a Concept Learner

To learn class expressions we need to build a model of the concept learner that we want to use. It can be either EvoLearner, CELOE or OCEL. Depending on the algorithm you chose there are different initialization parameters which you can check here. Let’s start by setting a quality function.

Quality metrics

There is a default quality function to evaluate the quality of the found expressions but different concept learners have different default quality function. Therefore, you may want to set it explicitly. There are the following quality function:F1 Score, Predictive Accuracy, Precision, and Recall. To use a quality function, first create an instance of its class:

from ontolearn.metrics import Accuracy

pred_acc = Accuracy()

In the following example we have built a model of OCEL and we have specified some of the parameters which can be set for OCEL.

(Optional) If you have target concepts that you want to ignore check how to ignore concepts.

Create a model

from ontolearn.concept_learner import OCEL

model = OCEL(knowledge_base=kb, 
              quality_func = pred_acc,
              max_runtime=600,  
              max_num_of_concepts_tested=10_000_000_000,  
              iter_bound=10_000_000_000)

The parameter knowledge_base which is the only required parameter, specifies the knowledge base that is used to learn and test concepts. The following parameters are optional.

  • quality_func - function to evaluate the quality of solution concepts. (Default value = F1())

  • max_runtime - runtime limit in seconds. (Default value = 5)

  • max_num_of_concepts_tested - limit to stop the algorithm after n concepts tested. (Default value = 10_000)

  • iter_bound - limit to stop the algorithm after n refinement steps are done. (Default value = 10_000)

Execute and fetch the results

After creating the model you can fit the learning problem into this model, and it will find the hypotheses that explain the positive and negative examples. You can do that by calling the method fit :

model.fit(lp)

The hypotheses can be saved:

model.save_best_hypothesis(n=3, path='Predictions')

save_best_hypothesis method creates a .owl file of the RDF/XML format containing the generated (learned) hypotheses. The number of hypotheses is specified by the parameter n. path parameter specifies the name of the file.

If you want to print the hypotheses you can use the method best_hypotheses which will return the n best hypotheses together with some insights such as quality value, length, tree length and tree depth of the hypotheses, and the number of individuals that each of them is covering, use the method best_hypotheses where n is the number of hypotheses you want to return.

hypotheses = model.best_hypotheses(n=3)  
[print(hypothesis) for hypothesis in hypotheses]

You can also create a binary classification for the specified individuals by using the predict method as below:

binary_classification = model.predict(individuals=list(typed_pos | typed_neg), hypotheses=hypotheses)

Here we are classifying the positives and negatives individuals using the generated hypotheses. This will return a data frame where 1 means True and 0 means False.

Verbalization

You can as well verbalize or visualize the generated hypotheses into images by using the static method verbalize. This functionality requires an external package which is not part of the required packages for Ontolearn as well as graphviz.

  1. Install deeponto. pip install deeponto + further requirements like JDK, etc. Check https://krr-oxford.github.io/DeepOnto/ for full instructions.

  2. Install graphviz at https://graphviz.org/download/.

After you are done with that you can simply verbalize predictions:

model.verbalize('Predictions.owl')

This will create for each class expression inside Predictions.owl a .png image that contain the tree representation of that class expression.


Use Triplestore as Knowledge Base

Instead of going through nodes using expensive computation resources why not just make use of the efficient approach of querying a triplestore using SPARQL queries. We have brought this functionality to Ontolearn for our learning algorithms, and we take care of the conversion part behind the scene. Let’s see what it takes to make use of it.

First of all you need a server which should host the triplestore for your ontology. If you don’t already have one, see Loading and Launching a Triplestore below.

Now you can simply initialize a TripleStore object that will server as an input for your desired concept learner as follows:

from ontolearn.triple_store import TripleStore

kb = TripleStore("http://your_domain/some_path/sparql")

Notice that the triplestore endpoint is the only argument that you need to pass. Also keep in mind that this knowledge base contains a TripleStoreOntology and TripleStoreReasoner which means that every querying process concerning concept learning is now using the triplestore.

Important notice: The performance of a concept learner may differentiate when using triplestore. This happens because some SPARQL queries may not yield the exact same results as the local querying methods.

Loading and Launching a Triplestore

We will provide a simple approach to load and launch a triplestore in a local server. For this, we will be using apache-jena and apache-jena-fuseki. As a prerequisite you need JDK 11 or higher and if you are on Windows, you need Cygwin. In case of issues or any further reference please visit the official page of Apache Jena and check the documentation under “Triple Store”.

Having that said, let us now load and launch a triplestore on the “Father” ontology:

Open a terminal window and make sure you are in the root directory. Create a directory to store the files for Fuseki server:

mkdir Fuseki && cd Fuseki

Install apache-jena and apache-jena-fuseki. We will use version 4.7.0.

# install Jena
wget https://archive.apache.org/dist/jena/binaries/apache-jena-4.7.0.tar.gz
#install Jena-Fuseki
wget https://archive.apache.org/dist/jena/binaries/apache-jena-fuseki-4.7.0.tar.gz

Unzip the files:

tar -xzf apache-jena-fuseki-4.7.0.tar.gz
tar -xzf apache-jena-4.7.0.tar.gz

Make a directory for our ‘father’ database inside jena-fuseki:

mkdir -p apache-jena-fuseki-4.7.0/databases/father/

Now just load the ‘father’ ontology using the following commands:

cd ..

Fuseki/apache-jena-4.7.0/bin/tdb2.tdbloader --loader=parallel --loc Fuseki/apache-jena-fuseki-4.7.0/databases/father/ KGs/Family/father.owl

Launch the server, and it will be waiting eagerly for your queries.

cd Fuseki/apache-jena-fuseki-4.7.0 

java -Xmx4G -jar fuseki-server.jar --tdb2 --loc=databases/father /father

Notice that we launched the database found in Fuseki/apache-jena-fuseki-4.7.0/databases/father to the path /father. By default, jena-fuseki runs on port 3030 so the full URL would be: http://localhost:3030/father. When you pass this url to triplestore_address argument, you have to add the /sparql sub-path indicating to the server that we are querying via SPARQL queries. Full path now should look like: http://localhost:3030/father/sparql.

You can now create a triplestore knowledge base or a reasoner that uses this URL for their operations: TODO