Correspondence pattern attribute selection for consumption of federated data sources

When consuming data from federated domains, it is often necessary to identify the relationships that exist between the data schemas used in each domain. Discovering the exact nature of these relationships is difficult due to data set schema heterogeneity. Prior work has focused on inter-domain class equivalence. However it is not always possible to find an equivalent class in both schemas. For example, when instances are modeled as classes in one domain (e.g. router type) but as the attribute values of a single class in the other domain (e.g. router interface). This paper investigates whether when classifying instances in one data set against a second schema, it may be more useful to use some attribute (or attribute group) other than the original class type, to perform this classification. A machine-learning based classification approach to appropriate attribute selection is presented and its operation is evaluated using two large data-sets available on the web as Linked Data. The classification problem is compounded by the less formal semantics of Linked Data when compared to full ontologies but this also highlights the strength of our approach to dealing with noisy or under-specified data-sets and schemas. The experimental results show that our attribute selection approach is capable of discovering appropriate mappings for cases where the correspondence is conditioned on one attribute and that information gain provides a suitable scoring function for selection of correspondence patterns to describe these complex attribute-based mappings.