Detecção de fraudes em licitações públicas através da identificação de anomalias nos valores cobrados e da análise de redes complexas formadas por interações de compra

Abstract

Anomaly detection in datasets is of great relevance to multiple current scenarios, particularly in the accounting field where anomalies can be considered signs of fraud. This type of analysis can be used to mitigate risks and avoid financial losses, especially in the public sector. The goal of this project is to use anomaly detection methods in invoice data from the Brazilian public sector autonomously, in order to help and speed up the work of financial auditors. Two distinct fronts were explored: (i) anomaly detection in structured data using Local Outlier Factor (LOF), Isolation Forest (iForest) and Self-Organizing Maps (SOM); and (ii) anomaly detection in complex networks using Generative Adversarial Attributed Network Anomaly Detection (GAAN) and Contrastive self-supervised Learning framework for Anomaly detection on attributed networks (CoLA). The results with higher anomaly scores were analyzed manually by the developers and, after comparing them with similar entries, were categorized according to their probability of being fraudulent instances. For the detection in structured data (i), the iForest proved to be the most promising method for detecting frauds, while the LOF showed unsatisfactory results; the SOM method proved to be more effective in detecting isolated outliers, making it impossible to identify fraud cases. For the detection in complex networks (ii), the CoLA model showed better results in the identification of irregular nodes where it was possible to observe differences between a node pointed by the model and other similar nodes, indicating that the node is, in fact, an anomaly instance. The GAAN model identified greatly isolated anomaly nodes, in other words nodes with few connections and few similar nodes, making it difficult to analyze the results and to identify if the instance is an anomaly or not.