Bayesian Networks for Matcher Composition in Automatic Schema Matching

Abstract

We propose a method for accurate combining of evidence supplied by multiple individual matchers regarding whether two data schema elements match (refer to the same object or concept), or not, in the field of automatic schema matching. The method uses a Bayesian network to model correctly the statistical correlations between the similarity values produced by individual matchers that use the same or similar information, in order to avoid overconfidence in match probability estimates and improve the accuracy of matching. Experimental results under several testing protocols suggest that the matching accuracy of the Bayesian composite matcher can significantly exceed that of the individual component matchers.