An Extendable Meta-learning Algorithm for Ontology Mapping

In this paper, we describe a machine learning approach to ontology mapping. Although Machine learning techniques have been used earlier in many semantic integration approaches, dependence on precision recall curves to preset the weights and thresholds of the learning systems has been a serious bottleneck. By recasting the mapping problem to a classification problem we try to automate this step and develop a robust and extendable meta learning algorithm. The implication is that we can now extend the same method to map the ontology pairs with different similarity measures which might not be specialized for the specific domain, yet obtain results comparable to the state of the art mapping algorithms that exploit machine learning methods. Interestingly we see that as the similarity measures are diluted, our approach performs significantly better for unbalanced classes. We have tested our approach using several similarity measures and two real world ontologies, and the test results we discuss validate our claim. We also present a discussion on the benefits of the proposed meta learning algorithm.