Automated matching of data mining dataset schemata to background knowledge

Interoperability in data mining is supported by a standard for dataset and model representation: Predictive Model Markup Language (PMML).1 It allows to describe the columns (which are continuous, categorical or ordinal) of the source data table, preprocessing transformations (such as discretization of continuous values) as well as the structure of the discovered model (e.g. neural network or set of association rules). In addition to source data, the input to the mining task typically includes expertprovided background knowledge. It may be related, for example, to standard ways of discretizing numerical quantities (e.g. boundaries between ‘normal blood pressure’ and ‘hypertension’, which help intuitive reading of discovered hypotheses), or may itself have the form of predictive models to which the discovered models can be compared during or after the mining process. The proposal for Background Knowledge Exchange Format (BKEF) [?], in many aspects similar to PMML, aims to support interoperability of data mining (and related) applications dealing with background knowledge. Case studies [?] showed that one BKEF model typically has to be aligned with different PMML models (from different mining sessions in the same domain). The alignments are stored in the Field Mapping Language (FML) [?], expressing that a data field (column) in a PMML model semantically corresponds to an abstract ‘field’ in a BKEF model. However, writing FML alignments (analogous to instances of the Alignment Format [?] used in ontology matching) by hand is tedious and recognizing suitable correspondences may be hard; partial automation is thus desirable. Furthermore, existing tools for ontology/schema matching are not straightforwardly usable, since PMML (and even BKEF) are, compared to ontologies, more biased by data structures, but BKEF is more abstract and weakly structured than database schemata. Therefore, specific methods (inspired by existing ones) and a new tool have been devised.