Learning-based fusion for data deduplication: A robust and automated solution.

摘要

This thesis presents two deduplication techniques that overcome the following critical and long-standing weaknesses of rule-based deduplication: (1) traditional rule-based deduplication requires significant manual tuning of the individual rules, including the selection of appropriate thresholds; (2) the accuracy of rule-based deduplication degrades when there are missing data values, significantly reducing the efficacy of the expert-defined deduplication rules.;The first technique is a novel rule-level match-score fusion algorithm that employs kernel-machine-based learning to discover the decision threshold for the overall system automatically. The second is a novel clue-level match-score fusion algorithm that addresses both Problem 1 and 2. This unique solution provides robustness against missing/incomplete record data via the selection of a best-fit support vector machine. Empirical evidence shows that the combination of these two novel solutions eliminates two critical long-standing problems in deduplication, providing accurate and robust results in a critical area of rule-based deduplication.