Unsupervised profiling for identifying superimposed fraud

Abstract

Many fraud analysis applications try to detect “probably fraudulent” usage patterns, and to discover these patterns in historical data. This paper builds on a different detection concept; there are no fixed “probably fraudulent” patterns, but any significant deviation from the normal behavior indicates a potential fraud. In order to detect such deviations, a comprehensive representation of “customer behavior” must be used. This paper presents such representation, and discusses issues derived from it: a distance function and a clustering algorithm for probability distributions.