Fraud/Uncollectible Debt Detection Using a Bayesian Network Based Learning System: A Rare Binary Outcome with Mixed Data Structures

TI1e fraud/uncollectible debt1 problem in the telecommunications industry presents two technical challenges: the detection and the treaunent of the account given the detection. In this paper, we focus on the first problem of detection using Bayeshm network models, and we briefly discuss U1e application of a nonnative expert system for U1e treatment at tl1e end. We apply Bayesian network models to the problem of fraud/uncollectible debt detection for telecommunication services. In addition to being quite successful at predicting rare event outcomes, it is able to handle a mixture of categorical and continuous data. We present a performance comparison using linear and non­ linear discriminant analysis, classification and regression trees, and Bayesian network models.