Towards a more Accurate and Fair SVM-based Record Linkage.

Record linkage, the process of identifying records representing the same real world entity in the absence of common unique identifiers, is one of the most intriguing problems in data processing, hence drawing attention for several decades. One approach for addressing this problem is by means of supervised learning. However, when it comes to linking records of individuals, the quality of the results may be low due to hidden bias phenomena, not deliberately caused, but originating from specific properties of the names of people of specific ethnic origin. In this paper, we focus on SVM-based record linkage and considering the fact of hidden bias, we propose a methodology oriented towards Ethnicity group-based training for specific Ethnicity groups, an approach that manages to elevate matching performance, compared to the average case.