It Pays to Be Certain: Unsupervised Record Linkage via Ambiguity Minimization

Record linkage (RL) is a process of identifying records that refer to the same real-world entity. Many existing approaches to RL apply supervised machine learning (ML) techniques to generate a classification model that classifies a pair of records as either linked or non-linked. In such techniques, the labeled data helps guide the choice and relative importance to similarity measures to be employed in RL. Unsupervised RL is therefore a more challenging problem since the quality of similarity measures needs to be estimated in the absence of linkage labels. In this paper we propose a novel optimization approach to unsupervised RL. We define a scoring technique which aggregates similarities between two records along all attributes and all available similarity measures using a weighted sum formulation. The core idea behind our method is embodied in an objective function representing the overall ambiguity of the scoring across a dataset. Our goal is to iteratively optimize the objective function to progressively refine estimates of the scoring weights in the direction of lesser overall ambiguity. We have evaluated our approach on multiple real world datasets which are commonly used in the RL community. Our experimental results show that our proposed approach outperforms state-of-the-art techniques, while being orders of magnitude faster.