Solving the False Negative Problem
By John Talburt, PhD, CDMP, Director, UALR Laboratory for Advanced Research in Entity Resolution and Information Quality (ERIQ)
In my March 25, 2009 post “The Myth of Matching,” I discussed the confusion between entity resolution and matching as in record de-duplication. Matching is a necessary part of entity resolution, but it is not sufficient. In particular I brought up the issue of “false negatives,” cases where records don’t match, but are in fact references to the same entity. I used the example of Mary Doe living on Elm Street who married John Smith living on Pine Street resulting in two references “Mary Doe, 234 Elm St” and “Mary Smith, 456 Pine St” that don’t match, but are never-the-less references to the same person. Let’s discuss a couple of approaches to solving this problem - enlarging the scope of identity attributes and utilizing asserted associations.
The Mary Doe - Mary Smith case might be resolved if the scope of identity attributes were increased, i.e. if additional information such as date-of-birth, drivers license, or social security number were available in both records. But as anyone acquainted with information quality understands, acquiring and maintaining additional information can create as many problems as it solves. It also brings up a number of questions that the information custodians and collectors must answer.
Is this information available? Is it costly? Is use for this purpose permissible/legal? Even if expanding the number of identity attributes is an option, it is not necessarily a panacea. Increasing the number of identity attributes also increases the complexity of the matching. What if some values are missing? What if some values agree, but others disagree?
A second approach is to collect and use asserted associations. The fundamental problem is that if Mary Doe and Mary Smith do not share any matching identity attributes, you cannot know that they are the same person without some separately acquired knowledge that they are in fact the same person. Moreover, because not all Mary Doe’s are the same person as Mary Smith, you also need additional context such as the address to make the connection clear. The upshot is that you need to possess the explicit knowledge that “Mary Doe at 234 Elm St is the same person as Mary Smith at 456 Pine St.”
If Mary lives in the United States and Mary registers her change of name and address with the US Postal Service, then you might be able to resolve this through the USPS Change of Address file. Besides the fact that this is only helpful in the US, relying on the USPS COA file has other disadvantages, not the least of which is that Mary may have decided not to register with the USPS. For this reason, some companies choose to maintain their own knowledge by acquiring information from other public and private sources.
For example in the US, marriage records are publicly available and are a possible source of this associative information. It may also be true that while Mary didn’t register her change of address with the USPS, she may have wanted to avoid missing any issues of her Modern Square Dancing magazine subscription and promptly registered her change of address with the publisher. There are potentially many other data sources, such as changes in utility service, cable service, or required licensure notifications.
Even though the application of external association information can alleviate the false negative problem, it comes at a cost. The collection and maintenance of associative information can be a monumental task for some types of entities. For example, at least 20% of the US population moves each year. Because it is too large a task for most organizations to take on by themselves, companies that aggregate large amounts of associative data sometimes offer the application of this knowledge as a product.
In the next installment, I will discuss another common confusion, the difference between entity resolution and identity resolution.
Subscribe to this blog