Record matching is the task of identifying records that match the same real world entity. This is a problem of great significance for a variety of business intelligence applications. Implementations of record matching rely on exact as well as approximate string matching (e.g., edit distances) and use of external reference data sources. Record matching can be viewed as a query composed of a small set of primitive operators. However, formulating such record matching queries is difficult and depends on the specific application scenario. Specifically, the number of options both in terms of string matching operations as well as the choice of external sources can be daunting. In this paper, we exploit the availability of positive and negative examples to search through this space and suggest an initial record matching query. Such queries can be subsequently modified by the programmer as needed. We ensure that the record matching queries our approach produces are (1) efficient: these queries can be run on large datasets by leveraging operations that are well-supported by RDBMSs, and (2) explainable: the queries are easy to understand so that they may be modified by the programmer with relative ease. We demonstrate the effectiveness of our approach on several real-world datasets.
展开▼
