Multicriteria Analysis of Clustering Algorithms for Anomaly Detection and Underbilling in Customs Operations

This article proposes a hybrid approach that integrates unsupervised machine learning and a multicriteria decision support method for customs fraud detection. Using real and anonymized cross-border e-commerce data from the Brazilian Federal Revenue Service, the performance of five clustering algorithms is evaluated using the AHP and PROMETHEE II methods. The evaluation contrasts internal mathematical validation metrics with the business rules of tax auditing. Validated by Stochastic Multicriteria Acceptability Analysis simulations, the results indicated the Gaussian Mixture Model as the superior alternative. The profiling of the anomalous cluster revealed a structural pattern of under-invoicing, optimizing the accuracy of tax screening and reducing the false alarm rate. It is concluded that the application of multicriteria decision support methods is highly effective in incorporating strategic and logistical perspectives into artificial intelligence models applied to public management.