Modelagem preditiva da estabilidade de formulações cosméticas multicomponentes: um estudo comparativo de modelos de aprendizado de máquina

Abstract

Physicochemical stability is a critical attribute in the development of multicomponent cosmetic formulations, and its assessment traditionally relies on time-consuming empirical tests. In this context, the present study aimed to compare the performance of tree-based machine learning models, such as Decision Tree, Random Forest and XGBoost, in the classification of shampoo formulation stability, using ingredient mass fractions as predictor variables. As a database, the experimental dataset of 812 formulations available in the literature was employed, from which stability was selected as the binary target variable (stable or unstable). The methodology involved data preparation and exploratory data analysis, followed by stratified splitting into train and test sets, cross-validation and hyperparameter optimization through grid search (GridSearchCV). The exploratory analysis revealed high sparsity, strong asymmetry and the absence of linear separability between classes, which justified the use of non-linear models. Performance was evaluated using F1-score, accuracy, precision and recall metrics, as well as error analysis through confusion matrices. A consistent progression in performance was observed with increasing model complexity: the Decision Tree, used as a reference, achieved an F1-score of 0.66 and accuracy of 68%, while the ensemble models, Random Forest and XGBoost, improved these metrics to an equivalent extent, reaching a test F1-score of 0.71 and accuracy of approximately 77%. Given the statistical equivalence between the two ensemble models on the test set, the mean cross-validation performance was adopted as the selection criterion, in which Random Forest achieved the best result (F1-score of 0.70), being selected as the final model. Feature importance analysis indicated thickeners and conditioning polymers as the most influential ingredients on stability, consistent with the combinatorial nature of the system. The results illustrate the potential of tree-based predictive models as a tool to support more agile and data-driven cosmetic formulation development.