Anomaly Detection for Fair Fraud Analysis (ADFA)
Fraud detection is essentially an anomaly detection procedure that aims to avoid financial loss. We propose a two step method to select cases probabilistically based on their impact, but guaranteeing that the relative discrepancy between the observed values and the expected behavior is also taken into account. Our method is multivariate, which allows it to consider several relevant variables simultaneously. The first step classifies observations based on the maximum distribution of a multivariate random vector. The second one uses a modification of the False Discovery Rate step-up Benjamini Hochberg procedure to improve accuracy and ensure scalability. We apply the proposed method to two different cases. The first one is a project designed to monitor the Brazilian public health care payment system in search for fraudulent behavior. The second is an analysis of the public finances of several municipalities in Brazil.