The Optimization of Hyperparameters in Unsupervised Learning Algorithms for Anomaly Detection in Public Procurement in Paraguay

Abstract

This study focuses on hyperparameter optimization in unsupervised learning algorithms for anomaly detection in public procurement processes in Paraguay. The main objective is to develop a tool that identifies irregularities in procurement processes using open data provided by the National Directorate of Public Procurement. The methodology follows the CRISP-DM industry standard, including data collection, transformation, and preparation, followed by the application of the algorithms Isolation Forest, Local Outlier Factor and One-Class SVM. Hyperparameter optimization is performed using grid search and random search techniques, and class imbalance is addressed using SMOTE oversampling. Results indicate that while the high recall model detects most anomalies, it produces a significant number of false positives. In contrast, to obtain models with high precision, a balancing of the data set is required, considerably reducing false positives at the cost of not identifying all anomalies. In conclusion, it is desirable to work on a correct labeling and balancing of the training data set to improve the accuracy and practical utility of the models.