John R Talburt

This paper discusses recent research on methods for estimating configuration parameters for the Matrix Comparator used for linking unstandardized or heterogeneously standardized references. The matrix comparator computes the aggregate similarity between the tokens (words) in a pair of references. The two most critical parameters for the matrix comparator for obtaining the best linking results are the value of the similarity threshold and the list of stop words to exclude from the comparison. Earlier research has shown that the standard deviation of the token frequency distribution is strongly predictive of how useful stop words will be in improving linking performance. The research results presented here demonstrate a method for using statistics from token frequency distribution to estimate the threshold value and stop word selection likely to give the best linking results. The model was made using linear regression and validated with independent datasets.

This paper describes a series of experiments in using logistic regression machine learning as a method for entity resolution. From these experiments the authors concluded that when a supervised ML algorithm is trained to classify a pair of entity references as linked or not linked pair, the evaluation of the model’s performance should take into account the transitive closure of its pairwise linking decisions, not just the pairwise classifications alone. Part of the problem is that the measures of precision and recall as calculated in data mining classification algorithms such as logistic regression is different from applying these measures to entity resolution (ER) results.. As a classifier, logistic regression precision and recall measure the algorithm’s pairwise decision performance. When applied to ER, precision and recall measure how accurately the set of input references were partitioned into subsets (clusters) referencing the same entity. When applied to datasets containing more than two references, ER is a two-step process. Step One is to classify pairs of records as linked or not linked. Step Two applies transitive closure to these linked pairs to find the maximally connected subsets (clusters) of equivalent references. The precision and recall of the final ER result will generally be different from the precision and recall measures of the pairwise classifier used to power the ER process. The experiments described in the paper were performed using a well-tested set of synthetic customer data for which the correct linking is known. The best F-measure of precision and recall for the final ER result was obtained by substantially increasing the threshold of the logistic regression pairwise classifier.