Bayesian Nonparametric Methods for Record linkage

Record linkage is a class of statistical and algorithmic methods which aim at identifying whether two or more observed records refer to the same statisti- cal entity or not. Duplications of the same entity within one single source or across different files may be interpreted as “clusters of records”, showing strong similar- ities across their fields. In this paper we frame the record linkage process into a formal Bayesian clustering model and we investigate the role of species sampling models as the natural prior specification for the clustering structure. In fact the dif- ferent latent entities which produce the records observed in one or more data sources can be effectively treated as the sampled species and the observed records as noisy measurements of their features. We also discuss an important issue in the cluster- ing approach to entity resolution, that is the necessity of bounding the clusters sizes even in the presence of large data sets.