Fraud detection in social income transfer programs: a social data mining approach applied to data from Brazil

Several assistance policies have been adopted by the Brazilian government to minimize social problems. As a basis for these actions, the government created the Unified Registry for Social Programs of the Federal Government (CadÚnico). The CadÚnico database, comprising more than 20 million records and 65 attributes, was constructed with the aim of storing information about every person at a social risk in Brazil. This study aims to identify possible fraudulent cases in Brazilian social policy claims involving cash transfer using a social data mining approach. The approach takes into account the experiences of social workers besides implementing traditional data mining techniques (e.g., decision trees, generalized linear models, BayesNets, support vector machines, etc.). Via the proposed method, we identified more than 25 thousand cases of possible fraud with a success rate of 98.69%. We utilized the knowledge of groups of specialists in urban, state, and national social policies, together with data mining techniques, for validation. Identification of such cases is expected to aid the formulation of an approach that can address social demand based on correct social data.