Unsupervised Evaluation of Entity Resolution

What is it about?

Entity resolution (ER) is the task of grouping together records that correspond to a single entity, such as a person or a product, in the absence of a unique identifier. For example, assume there is a record of a John Smith from Queensland, Australia, in a health database, and another record of a John Smith from Melbourne, Australia in an immigration database. If there is no unique personal identifier common to these databases (such as a Tax File Number, which is unique to a single individual), we need to rely on ER techniques to determine whether these two records refer to one single person (true match) who has potentially moved from Queensland to Melbourne, or two different people (true non match) with the same name. A main challenge in ER is the lack of ground truth data in the form of known true matches and non matches. This hinders the ability to assess how well ER techniques perform in real world applications. In this paper, we propose novel methods to estimate the quality of ER techniques in the absence of ground truth data, which are therefore referred to as unsupervised evaluation techniques of ER.

Featured Image

Photo by Growtika on Unsplash

Photo by Growtika on Unsplash

Why is it important?

Many organisations that apply ER for operational purposes struggle to determine how confident they can be about the accuracy of resolved entities due to the absence or lack of ground truth data. Our proposed quality estimation approaches are helpful in these real-world scenarios to gain an idea of how good a given ER technique may or may not be.

Perspectives