Leveraging feature selection to detect potential tax fraudsters

Tax evasion is any act that knowingly or unknowingly, legally or unlawfully, leads to non-payment or underpayment of tax due. Enforcing the correct payment of taxes by taxpayers is fundamental in maintaining investments that are necessary and benefits a society as a whole. Indeed, without taxes it is not possible to guarantee basic services such as health-care, education, sanitation, transportation, infrastructure, among other services essential to the population. This issue is especially relevant in developing countries such as Brazil. In this work we consider a real-world case study involving the Treasury Office of the State of Ceará (SEFAZ-CE, Brazil), the agency in charge of supervising more than 300,000 active taxpayers companies. SEFAZ-CE maintains a very large database containing vast amounts of information concerning such companies. Its enforcement team struggles to perform thorough inspections on taxpayers accounts as the underlying traditional human-based inspection processes involve the evaluation of countless fraud indicators (i.e., binary features), thus requiring burdensome amounts of time and being potentially prone to human errors. On the other hand, the vast amount of taxpayer information collected by fiscal agencies opens up the possibility of devising novel techniques able to tackle fiscal evasion much more effectively than traditional approaches. In this work we address the problem of using feature selection to select the most relevant binary features to improve the classification of potential tax fraudsters. Finding out possible fraudsters from taxpayer data with binary features presents several challenges. First, taxpayer data typically have features with low linear correlation between themselves. Also, tax frauds may originate from intricate illicit tactics, which in turn requires to uncover non-linear relationships between multiple features. Finally, few features may be correlated with the targeted class. In this work we propose Alicia, a new feature selection method based on association rules and propositional logic with a carefully crafted graph centrality measure that attempts to tackle the above challenges while, at the same time, being agnostic to specific classification techniques. Alicia is structured in three phases: first, it generates a set of relevant association rules from a set of fraud indicators (features). Subsequently, from such association rules Alicia builds a graph, which structure is then used to determine the most relevant features. To achieve this Alicia applies a novel centrality measure we call the Feature Topological Importance. We perform an extensive experimental evaluation to assess the validity of our proposal on four different real-world datasets, where we compare our solution with eight other feature selection methods. The results show that Alicia achieves F-measure scores up to 76.88%, and consistently outperforms its competitors.