Sammon Mapping-Based Gradient Boosted Trees for Tax Crime Prediction in the City of São Paulo

With the currently vast volume of data available, several institutions, including the public sector, benefit from information, aiming to improve decision-making. Machine Learning enhances data-driven decision-making with its predictive power. In this work, our principal motivation was to apply Machine Learning to ameliorate fiscal audit planning for São Paulo’s municipality. In this study, we predicted crimes against the service tax system of São Paulo using Machine Learning. Our methodology embraced the following steps: data extraction; data preparation; dimensionality reduction; model training and testing; model evaluation; model selection. Our experimental findings revealed that Sammon Mapping (SM) combined with Gradient Boosted Trees (GBT) outranked other state-of-the-art works, classifiers and dimensionality reduction techniques as regards classification performance. Our belief is that the ensemble of classifiers of GBT, combined with SM’s ability to identify relevant dimensions in data, contributed to produce higher prediction scores. These scores enable São Paulo’s tax administration to rank fiscal audits according to the highest probabilities of tax crime occurrence, leveraging tax revenue.