Evaluate a class expression

When using a concept learner, the generated concepts (class expressions) for a certain learning problem can be evaluated using different approaches. One of them is to use the static functions evaluate_concept. It requires the following arguments:

  1. a knowledge base: AbstractKnowledgeBase

  2. a concept to evaluate: OWLClassExpression

  3. a quality function: AbstractScorer

  4. the encoded learning problem: EncodedLearningProblem

When you use a concept learner to generate a class expression (CE) you usually evaluate the performance of the CE in the learning problem examples that you used to generate the CE in the first place. The function evaluate_concept returns an object of EvaluatedConcept.

In this guide we will not use a concept learner to generate a CE but will construct a learning problem and a CE manually. Furthermore, we will consider you already have created an object of AbstractKnowledgeBase (either KnowlegeBase or TripleStore). If that’s not the case, check Knowledge Bases.

Construct a learning problem

Let’s define a learning problem for the Father ontology. Firstly, we create two simple sets containing the positive and negative examples for the concept of ‘Father’. Our positive examples (individuals to describe) are stefan, markus, and martin. And our negative examples (individuals to not describe) are heinz, anna, and michelle.

from owlapy.owl_individual import OWLNamedIndividual, IRI

positive_examples = {OWLNamedIndividual(IRI.create(NS, 'stefan')),
                     OWLNamedIndividual(IRI.create(NS, 'markus')),
                     OWLNamedIndividual(IRI.create(NS, 'martin'))}

negative_examples = {OWLNamedIndividual(IRI.create(NS, 'heinz')),
                     OWLNamedIndividual(IRI.create(NS, 'anna')),
                     OWLNamedIndividual(IRI.create(NS, 'michelle'))}

Now the learning problem can be represented by using the positive-negative standard learning problem.

from ontolearn.learning_problem import PosNegLPStandard

lp = PosNegLPStandard(pos=positive_examples, neg=negative_examples)

We then encod it to the knowledge base:

encoded_lp = lp.encode_kb(kb)

where kb is an instance of KnowledgeBase or TripleStore, instantiated in our previous guide.

Now that we have an encoded learning problem, we need a concept to evaluate.

Construct a concept

Suppose that the class expression (¬female) ⊓ (∃ hasChild.⊤) was generated by CELOE for the target concept of ‘Father’. We will see how that can happen later but for now we let’s construct this class expression manually using owlapy:

from owlapy.owl_property import OWLObjectProperty
from owlapy.class_expression import OWLClass, OWLObjectSomeValuesFrom , OWLObjectIntersectionOf
from ontolearn.concept_generator import ConceptGenerator

generator = ConceptGenerator()

female = OWLClass(IRI(NS,'female'))
not_female = generator.negation(female)
has_child_property = OWLObjectProperty(IRI(NS, "hasChild"))
thing = OWLClass(IRI('http://www.w3.org/2002/07/owl#', 'Thing'))
exist_has_child_T = OWLObjectSomeValuesFrom(property=has_child_property, filler=thing)

concept_to_test = OWLObjectIntersectionOf([not_female, exist_has_child_T])

We use an instance of ConceptGenerator to create the negated concept ¬female. OWLClass represent a named class in OWL, OWLObjectProperty, represents an object property in OWL and OWLObjectSomeValuesFrom, OWLObjectIntersectionOf represent class expression constructs.

Evaluation and results

You can now evaluate the class expression you just created:

from ontolearn.quality_funcs import evaluate_concept

evaluated_concept = evaluate_concept(concept_to_test, F1(), encoded_lp)

In this example we use F1-score to evaluate the concept, but there are more metrics which you can use including Accuracy, Precision and Recall.

You can now:

  • Print the quality:

    print(evaluated_concept.q) # 1.0

  • Print the set of individuals covered by the class expression:

    for ind in evaluated_concept.inds:
    print(ind) 

# OWLNamedIndividual(http://example.com/father#markus)
# OWLNamedIndividual(http://example.com/father#martin)
# OWLNamedIndividual(http://example.com/father#stefan)

  • Print the amount of them:

    print(evaluated_concept.ic) # 3

In the next guide we will walk through the steps of using a concept learning model.