METODOLOGIA DE UTILIZAÇÃO DO FRAMEWORK OPTUNA PARA OTIMIZAÇÃO DE HIPERPARÂMETROS DE MODELO DE APRENDIZADO DE MÁQUINA DE REGRESSÃO
This study presents a methodology for using the Optuna framework to optimize hyperparameters in regression machine learning models, illustrated by its application to XGBoost. The objective was to organize the analysis of optimization logs to support the definition of new search intervals. Eleven public datasets from the UC Irvine Machine Learning Repository were used to define fifteen tasks, with data preparation, categorical encoding, and normalization. The procedure, named MOHOC, combines Optuna trial logs with Microsoft Excel visualizations, enabling the analyst to examine concentrations of results and redefine hyperparameter intervals. Four optimization cycles were performed for each task, with 1500 trials in the first cycle. At the end, the root mean squared error (RMSE) decreased by an average of 15.91% relative to the default configurations and 4.24% relative to the first cycle. These results describe the observed search progression; they do not provide an independent assessment of generalization or isolate the effect of manual refinement from the increased optimization budget. The contribution is a practical procedure for analyzing and guiding successive search cycles.
Authors
- Fabricio Ribeiro Seppe (ORCID: https://orcid.org/0000-0002-4702-8823)
Institutions
- Escola Superior de Propaganda e Marketing (BR)
Publication Details
- Journal
- RECIMA21 - Revista Científica Multidisciplinar
- Published
- 2026-11-21
- DOI
- https://doi.org/10.47820/recima21.v7i12.9477
- Primary Topic
- Machine Learning and Data Classification
- Type
- article
- Field-Weighted Citation Impact
- 0.00