SERIMI: Class-based Disambiguation for Effective Instance Matching over Heterogeneous Web Data.

Abstract

Instance matching has been studied with focus on the singledomain setting, while less attention is given to the heterogeneous environment of the Web, where data comes from different domains and are associated with different schemas. For this heterogeneous setting, we propose an unsupervised schema-agnostic approach that focuses on the refinement (disambiguation) of candidate instances (resulting from blocking). Given instances of a source dataset that belong to a class, it computes candidates in the target datasets and refines them such that the remaining matches correspond to the source instances at the class level. However, no schema knowledge and explicit correspondences between classes in the source and target datasets are required for this. Rather, the disambiguation is performed based on an instance-based representation of classes computed online. We evaluated our work using experiments on large-scale real-world datasets provided by a benchmark. The proposed solution outperformed two alternative approaches for instance matching in 70 % of the cases, and in those cases we improved average F-measure by 10%.