Knowledge discovery from data?

(KDD) field draws on findings from statistics, databases, and artificial intelligence to construct tools that let users gain insight from massive data sets. People in business, science, medicine, academia, and government collect such data sets, and several commercial packages now offer general-purpose KDD tools. An important KDD goal is to “turn data into knowledge.” For example, knowledge acquired through such methods on a medical database could be published in a medical journal. Knowledge acquired from analyzing a financial or marketing database could revise business practice and influence a management school’s curriculum. In addition, some US laws require reasons for rejecting a loan application, which knowledge from the KDD could provide. Occasionally, however, you must explain the learned decision criteria to a court, as in the recent lawsuit Blue Mountain filed against Microsoft for a mail filter that classified electronic greeting cards as spam mail.1 In one early KDD success story, Robert Evans and Doug Fisher analyzed data from a printing press, found conditions under which the press failed, and identified rules to avoid these failures.2 Unfortunately, for every insightful nugget we find, there are many more obvious or trivial rules (such as “unemployed people don’t earn income from work”3). Perhaps more troubling is that some rules are counterintuitive. For example, in screening for Alzheimer’s disease, we found the following counterintuitive rule: “If the years of education of the patient is greater than 5 and the patient does not know the date and the patient does not know the name of a nearby street, then the patient is normal.”4