Résumé en anglais

Data integration typically seeks to provide the illusion that data from multiple distributed sources comes from a single, well managed source. Providing this illusion in practice tends to involve the design of a global schema that captures the users data requirements, followed by manual (with tool support) construction of mappings between sources and the global schema. This overall approach can provide high quality integrations but at high cost, and tends to be unsuitable for areas with large numbers of rapidly changing sources, where users may be willing to cope with a less than perfect integration. Pay-as-you-go data integration has been proposed to overcome the need for costly manual data integration. Pay-as-you-go data integration tends to involve two steps. Initialisation: automatic creation of mappings (generally of poor quality) between sources. Improvement: the obtaining of feedback on some aspect of the integration, and the application of this feedback to revise the integration. There has been considerable research in this area over a ten year period. This paper reviews some experiences with pay-as-you-go data integration, providing a framework that can be used to compare or develop pay-as-you-go data integration techniques.