Scalable Similarity Joins for Fast and Accurate Record Deduplication in Big Data

Abstract

Record linkage is the process of matching records from multiple data sources that refer to the same entities. When applied to a single data source, this process is known as deduplication. With the increasing size of data source, recently referred to as big data, the complexity of the matching process becomes one of the major challenges for record linkage and deduplication. In recent decades, several blocking, indexing and filtering techniques have been developed. Their purpose is to reduce the number of record pairs to be compared by removing obvious non-matching pairs in the deduplication process, while maintaining high quality of matching. Currently developed algorithms and traditional techniques are not efficient, using methods that still lose significant proportion of true matches when removing comparison pairs. This paper proposes more efficient algorithms for removing non-matching pairs, with an explicitly proven mathematical lower bound on recently used state-of-the-art approximate string matching method - Fuzzy Jaccard Similarity. The algorithm is also much more efficient in classification using Density-based spatial clustering of applications with noise (DBSCAN) in log-linear time complexity $\mathcal {O}(|\mathcal {E}|\log (|\mathcal {E}|))$.