A discussion of some joint research with folks at the University of Texas on fraud detection via a binary classification of (insurance claim) characteristic vectors in n-space. This result fits into a "data mining" slot known as "unsupervised" learning, i.e. there are no known assignments to the two classes (fraud/ no fraud) but rather known or assumed responses (vector components) that are monotone in a latent variable (fraud/ no fraud). The origins of the technique are in educational testing (marketing) where the feature vectors are scored answers to questions and the latent variable is pass/fail (buy/no buy). Comparisons with other common modelling results for fraud and an application to structural changes in databases will be covered. No prior knowledge of insurance will be assumed or needed.
