Evaluation of approximate comparison methods on Bloom filters for probabilistic linkage

The need for increased privacy protection in data linkage has driven the development of privacy-preserving record linkage (PPRL) techniques. A popular technique using Bloom filters with cryptographic analyses, modifications, and hashing variations to optimise privacy has been the focus of much research in this area. With few applications of Bloom filters within a probabilistic framework, there is limited information on whether approximate matches between Bloom filtered fields can improve linkage quality.

In this study, we evaluate the effectiveness of three approximate comparison methods for Bloom filters within the context of the Fellegi-Sunter model of recording linkage: Sørensen-Dice coefficient, Jaccard similarity and Hamming distance.

Using synthetic datasets with introduced errors to simulate datasets with a range of data quality and a large real-world administrative health dataset, the research estimated partial weight curves for converting similarity scores (for each approximate comparison method) to partial weights at both field and dataset level. Deduplication linkages were run on each dataset using these partial weight curves. This was to compare the resulting quality of the approximate comparison techniques with linkages using simple cut-off similarity values and only exact matching.

Linkages using approximate comparisons produced significantly better quality results than those using exact comparisons only. Field level partial weight curves for a specific dataset produced the best quality results. The Sørensen-Dice coefficient and Jaccard similarity produced the most consistent results across a spectrum of synthetic and real-world datasets.

The use of Bloom filter similarity comparisons for probabilistic record linkage can produce linkage quality results which are comparable to Jaro-Winkler string similarities with unencrypted linkages. Probabilistic linkages using Bloom filters benefit significantly from the use of similarity comparisons, with partial weight curves producing the best results, even when not optimised for that particular dataset.