Fairness-aware Privacy-Preserving Record Linkage

Record linkage aims to identify records from different databases that correspond to the same real-world entity, while Privacy-Preserving Record Linkage (PPRL) conducts the linkage in a privacy-preserving context where private and sensitive information about individuals is not compromised. Linking records is considered as a classification task where pairs of records from different databases are classified into matches (i.e. they refer to the same entity) or non-matches (i.e. they refer to different entities). Due to the absence of unique entity identifiers across databases, commonly available quasi-identifiers (QIDs), such as name, gender, address, and date of birth, are used to determine the linkage. The values in such QIDs are often prone to data errors and variations making the linkage task challenging. Fairness in classification is an emerging concept that determines how much a classifier distorts from producing correct predictions with equal probabilities for individuals across different protected groups based on sensitive features (e.g. gender or race). Developing classifiers that are fair with respect to such sensitive features is an important problem for classification in general and specifically for PPRL to mitigate the bias against sensitive and/or minority groups, for example against female group due to higher likelihood of variations in the QIDs such as last name and address. While there have been increased interest in this field, fairness specifically in PPRL research has not been studied in the literature so far. Fairness for PPRL brings in specific challenges and requirements. In this paper, we study fairness for PPRL classifiers, analyse appropriate fairness criteria/metric for PPRL, study different forms of fairness-bias for PPRL and investigate the effectiveness of using fairness-aware PPRL. Our experimental results conducted on real and synthetically biased datasets show the efficacy and significance of incorporating fairness constraints in the linkage, leading to higher linkage quality in terms of both correctness and fairness.