Classification for Fraud Detection with Social Network Analysis
Miguel Pironet San-Bento Almeida · 2009
Worldwide fraud conducts to big losses to states’ treasuries and to private companies. Because of that, motivations to detect and fight fraud are high, but despite continuous efforts, it is far from being accomplished. The problems faced when trying to characterize fraud activities are many, with the specificities of fraud on each business domain leading the list. Despite the differences, building a classifier for fraud detection almost always requires to deal with unbalanced datasets, since fraudulent records are usually in a small number when compared with the nonfraudulent ones. This work describes two types of techniques to deal with fraud detection: techniques at a preprocessing level where the goal is to balance the dataset, and techniques at a processing level where the objective is to apply different errors costs to fraudulent and non-fraudulent cases. Besides that, as organizations and people more often do associations in order to commit fraud, is proposed a new method to make use of that information to improve the training of classifiers for fraud detection. In particular, this new method identifies patterns among the social networks for fraudulent organizations, and uses them to enrich the description of its entity. The enriched data will then be used jointly with balancing techniques to produce a better classifier to identify fraud.