Improving Search-based OWL Class Expression Learners

Topic

Given positive and negative examples, OWL class expression learners search for a class expression that describes the positive examples and excludes the negative ones. Search-based learners such as CELOE and EvoLearner are inherently explainable, as the learned expression is a first-order logical expression. EvoLearner is an evolutionary algorithm that evolves a population of class expressions through crossover and mutation, guided by a fitness function.

Yet, the search remains expensive on large knowledge graphs, and the quality of the learned expressions depends strongly on the initial population, the variation operators, and the fitness function.

In this thesis, the student will investigate how to improve an existing search-based learner such as EvoLearner. Possible directions include (but are not limited to)

• better initialization of the population,
• new crossover and mutation operators,
• fitness functions that balance predictive quality and expression length,
• techniques that preserve diversity in the population to avoid premature convergence,
• using knowledge graph embeddings (e.g. trained with dice-embeddings) to guide the search.

The proposed improvements will be evaluated against CELOE, EvoLearner, and other learners on benchmark datasets (e.g. Family, Mutagenesis, and Carcinogenesis) in terms of F1 score, expression length, and runtime.

The student will closely work on ontolearn and owlapy.

Question & Answer Session

In case you have further questions, feel free to contact Caglar Demir.