Learning Expressive Linkage Rules for Entity Matching using Genetic Programming
Robert Isele · MADOC (University of Mannheim) · 2013
A central problem in data integration and data cleansing is to identify pairs of entities in data sets that describe the same real-world object. Many existing methods for matching entities rely on explicit linkage rules, which specify how two entities are compared for equivalence. Unfortunately, writing accurate linkage rules by hand is a non-trivial problem that requires detailed knowledge of the involved data sets. Another important issue is the efficient execution of linkage rules. In this thesis, we propose a set of novel methods that cover the complete entity matching workflow from the generation of linkage rules using genetic programming algorithms to their efficient execution on distributed systems. First, we propose a supervised learning algorithm that is capable of generating linkage rules from a gold standard consisting of set of entity pairs that have been labeled as duplicates or non-duplicates. We show that the introduced algorithm outperforms previously proposed entity matching approaches including the state-of-the-art genetic programming approach by de Carvalho et al. and is capable of learning linkage rules that achieve a similar accuracy than the human written rule for the same problem. In order to also cover use cases for which no gold standard is available, we propose a complementary active learning algorithm that generates a gold standard interactively by asking the user to confirm or decline the equivalence of a small number of entity pairs. In the experimental evaluation, labeling at most 50 link candidates was necessary in order to match the performance that is achieved by the supervised GenLink algorithm on the entire gold standard. Finally, we propose an efficient execution workflow that can be run on cluster of multiple machines. The execution workflow employs a novel multidimensional indexing method that allows the efficient execution of learned linkage rules by reducing the number of required comparisons significantly.