Automatic schema merging using mapping constraints among incomplete sources

Abstract

Schema merging is the process of consolidating multiple schemas into a unified view. The task becomes particularly challenging when the schemas are highly heterogeneous and autonomous. Classical data integration systems rely on a mediated schema created by human experts through an intensive design process.In this paper, we present a novel approach for merging multiple relational data sources related by a collection of mapping constraints in the form of P2P style tuple-generating dependencies (tgds). In the scenario of data integration, we opt for minimal mediated schemas that are complete regarding certain answers of conjunctive queries. Under Open World Assumption (OWA), we characterize the semantics of schema merging by properties of the output mapping system between the source schemas and the mediated schema. We propose a merging algorithm following a redundancy reduction paradigm and prove that the output satisfies the desired logical properties. Recognizing the fact that multiple plausible mediated schemas may co-exist, a variant of the a priori algorithm is employed to enumerate alternative mediated schemas. Output mappings in the form of data dependencies are generated to support the mediated schemas, which enables query processing. We have evaluated our merging approach over a collection of real world data sets, which demonstrate the applicability and effectiveness of our approach in practice.

Merging Relational Views: A Minimization Approach

“…That is, there exists a backward mapping in full tgds recovering the unprojected database. We refer interested readers to [12] for more details.…”

“…We have described in [12] a procedure to test redundancy of a set of columns on the source schema with constraints, which is based on testing query rewritability of identity queries of source relations over the projected schema. The procedure can be extended to test redundancy of columns in a collapsed schema.…”

“…As clarified in [16], there are two semantically distinct flavors of schema merging, view integration [7,6,22,14,15,3] and data integration [17,8,19,20,12]. From a model theoretic perspective, a main distinction is that the former flavor assumes the input schemas are views of a base schema and hence satisfy constraints on the instance level, while the latter need to integrate autonomous input schemas which do not necessarily conform to any instance level constraints.…”

“…[16] also analyzed the different operational goals of the two types of schema merging. We study in this paper view integration, while our previous work [12] deals with data integration.…”

“…The formalism is extended in this paper to incorporate considerations on the size of the mediated schema, resulting in a characterization of Minimal Merge. We presented in [12] a merging algorithm that removed redundant columns from the source schemas in a data integration context. The approach described here considers, in addition, collapsing of relations, which leads to an even smaller schema.…”