An accurate tax fraud classifier with feature selection based on complex network node centrality measure

Fiscal evasion represents a very serious issue in many developing countries. In this context, tax fraud detection constitutes a challenging problem, since fraudsters change frequently their behaviors to circumvent existing laws and devise new kinds of frauds. Detecting such changes proves to be challenging, since traditional classifiers fail to select features that exhibit frequent variations. In this paper we provide two contributions that try to tackle effectively the tax fraud detection problem: first, we introduce a novel feature selection algorithm, based on complex network techniques, that is able to capture key fraud indicators - over time, this kind of indicators turn out to be more stable than new fraud indicators. Secondly, we propose a classifier that leverages the aforementioned algorithm to accurately detect tax frauds. In order to prove the validity of our contributions we provide an experimental evaluation, where we use real-world datasets, obtained from the State Treasury Office of Ceará (SEFAZ-CE), Brazil, to show how our method is able to outperform, in terms of F1 scores achieved, the state-of-the-art available in the literature.