Matcher composition methods for automatic schema matching

Abstract

We address the problem of automating the process of deciding whether two data schema elements match (that is, refer to the same actual object or concept), and propose several methods for combining evidence computed by multiple basic matchers. One class of methods uses Bayesian networks to account for the conditional dependency between the similarity values produced by individual matchers that use the same or similar information, so as to avoid overconfidence in match probability estimates and improve the accuracy of matching. Another class of methods relies on optimization switches that mitigate this dependency in a domain-independent manner. Experimental results under several testing protocols suggest that the matching accuracy of the Bayesian composite matchers can significantly exceed that of the individual component matchers, and the careful selection of optimization switches can improve matching accuracy even further.