Approximate String Matching by End-Users using Active Learning

Identifying approximately identical strings is key for many data cleaning and data integration processes, including similarity join and record matching. The accuracy of such tasks crucially depends on appropriate choices of string similarity measures and thresholds for the particular dataset. Manual selection of similarity measures and thresholds is infeasible. Other approaches rely on the existence of adequate historic ground-truth or massive manual effort. To address this problem, we propose an Active Learning algorithm which selects a best performing similarity measure in a given set while optimizing a decision threshold. Active Learning minimizes the number of user queries needed to arrive at an appropriate classifier. Queries require only the label match/no match, which end users can easily provide in their domain. Evaluation on well-known string matching benchmark data sets shows that our approach achieves highly accurate results with a small amount of manual labeling required.