ABSTRACT
142Fraud detection systems are a hot topic in the area of business intelligence. Traditionally, this field has been focused mostly on predictive analytics based on neural networks, Bayesian networks, logistic regression, and similar methods. Predictive analytics as a key driver for fraud detection modeling rarely delivers sufficient accuracy and can be significantly improved. Fortunately, fraud is not a frequent event, and it is always a problem to isolate a reliable data sample for a predictive analytical model. On the other hand, after conducting the attribute relevance analysis, predictive models recognize only a few important predictors regarding aim variable (fraud “yes” or “no”), which can be an insufficient tool for reaching adequate accuracy in fraud detection because fraudsters are very inventive. The predictive model is often based on a few integrated variables and is often a weak tool for fraud detection. Also, predictive model development depends on reliable history data samples. Reliability in fraud detection means a significant number of cases with common characteristics. Moreover, fraud demands quick recognition at an early stage, and pattern recognition in a late phase often becomes too slow and unusable for fraud prevention purposes. The authors do not neglect the importance of predictive models in fraud detection but emphasize that a more sophisticated approach should be used with significant improvement in accuracy. The main hypothesis is that qualitative fraud detection solutions should contain expert rules based on expert experience and should take into consideration other approaches, such as social network analysis, as well as Big Data sources for building efficient fraud detection models. The proposed approach integrates predictive models, fuzzy expert systems, social network analysis, and unstructured data, taking into account structured data from databases and data warehouses, social network data, and unstructured data from Internet resources within one solution. The chapter describes case studies in the area of fraud detection, based on proposed methodology. The case studies confirm the efficiency of the proposed methodology as a complex business intelligence system, which includes a traditional approach toward data sources as well as features of Big Data sources.
