• Fuzzy OWL
• Learning

This work aims at defining, implementing and comparing some algorithms that automatically learn specific OWL 2 inclusion axioms describing a target OWL 2 concept. To do so, a crisp OWL 2 background ontology, a target concept and a crisp training set are given. The learnt inclusion axioms belong to the fuzzy OWL 2 EL Profile language. Fuzziness has been used in order to improve the readability of the induced axioms (e.g., a good hotel is one that has a low price).

The algorithms taken under consideration are FOIL, a probabilistic variant of FOIL (pFOIL), a genetic variant of FOIL (gFOIL), the combination of the latter two (pgFOIL) and an AdaBoost variant of gFOIL (gAdaBoost).

FOIL learns one axiom trying to greedily maximizing a score function. After having learnt one axiom the positives samples covered are removed. The procedure is iterated until a give coverage threshold is reached.

pFOIL tries to learn axioms by taking into account the ensemble of learnt axioms, by evaluating the ensemble in probabilistic terms.

gFOIL exploits hybrid learning to learn one axiom. Hybrid learning is a particular form of genetic programming that here is improved with an ontology background theory. Like FOIL, the algorithm evaluates one axiom at time and removes the positive samples covered.

pgFOIL extends gFOIL by evaluating the goodness of a learnt axiom like for pFOIL.

gAdaBoost is based on a modified version of Real AdaBoost in which the weak learner used is as for hybrid learning.

Finally, a K-Fold cross validation procedure has been adopted to evaluate the learning algorithms.