Análise de performance entre TPOT e FLAML em problemas de classificação supervisionada

Abstract

The growing adoption of machine learning techniques in fields such as healthcare, finance, and marketing has driven the demand for solutions that automate the process of building predictive models. In this context, AutoML tools have gained prominence by enabling users with limited technical expertise to develop effective models with minimal manual intervention. This study aims to compare the predictive performance of the AutoML libraries TPOT and FLAML in binary classification problems. Three datasets with varying levels of class imbalance were used, applying a standardized preprocessing methodology, including imputation of missing values, categorical variable encoding, and data standardization. The experiments were conducted in a controlled environment, and the libraries were evaluated based on the ROC-AUC metric as well as execu- tion time. The results indicate that both tools achieved similar performance in terms of ROC-AUC, but FLAML showed a significant advantage in execution time, completing tasks that took TPOT several hours to process. Further analysis of the results revealed differences in optimization strategies and algorithm selection between the libraries. Based on the findings, it is concluded that FLAML is more suitable in contexts that demand higher computational efficiency, while TPOT may be preferred in scenarios that benefit from a broader search for models.