Learning to deduplicate

Identifying record replicas in Digital Libraries and other types of digital repositories is fundamental to improve the quality of their content and services as well as to yield eventual sharing efforts. Several deduplication strategies are available, but most of them rely on manually chosen settings to combine evidence used to identify records as being replicas. In this paper, we present the results of experiments we have carried out with a novel Machine Learning approach we have proposed for the deduplication problem. This approach, based on Genetic Programming (GP), is able to automatically generate similarity functions to identify record replicas in a given repository. The generated similarity functions properly combine and weight the best evidence available among the record fields in order to tell when two distinct records represent the same real-world entity. The results of the experiments show that our approach outperforms the baseline method by Fellegi and Sunter by more than 12% when identifying replicas in a data set containing researcher's personal data, and by more than 7%, in a data set with article citation data.