Paper 2

Entity resolution, which seeks to identify records that represent the same entity, is an important step in many data integration and data cleaning applications. However, entity resolution is challenging both in terms of scalability (all-against-all comparisons are computationally impractical) and result quality (syntactic evidence on record equivalence is often equivocal). As a result, end-to-end entity resolution proposals involve several stages, including blocking to efficiently identify candidate duplicates, detailed comparison to refine the conclusions from blocking, and clustering to identify the sets of records that may represent the same entity. However, the quality of the result is often crucially dependent on configuration parameters in all of these stages, for which it may be difficult for a human expert to provide suitable values. This paper describes an approach in which a complete entity resolution process is optimized, on the basis of feedback (such as might be obtained from crowds) on candidate duplicates. Given such feedback, an evolutionary search of the space of configuration parameters is carried out, with a view to maximizing the fitness of the resulting clusters. The approach is pay-as-you-go in that more feedback can be expected to give rise to better outcomes. An empirical evaluation shows that the co-optimization of the different stages in entity resolution can yield significant improvements over default parameters, even with small amounts of feedback.