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.
展开▼