Métodos de ciência de dados em People Analytics para previsão de turnover

Abstract

This research aimed to evaluate and compare the performance of different computational models applied to three public employee datasets for turnover prediction, considering class imbalance treatment strategies, different feature representations, and analyses stratified by sex. Six classifiers, one condition without resampling, and the SMOTE, ADASYN, SMOTE-Tomek, and SMOTE-ENN techniques were evaluated. For each population and experimental representation, 30 stratified 80% training and 20% testing splits were performed with different seeds, while transformations, Scenario 2 feature selection, and class balancing were restricted to the training set. The experiment comprised 24,300 model runs. Performance was assessed through accuracy, precision, recall, F1-score, MCC, AUC-ROC, and Log Loss, and the Weighted Multivariate Index was used to rank configurations within each context and examine their relative stability across repetitions. For the complete populations in Scenario 1, the reference configurations were LDA under the Imbalanced condition for Dataset 1, XGBoost with SMOTE-Tomek for Dataset 2, and AdaBoost with SMOTE-ENN for Dataset 3. Among these configurations, Dataset 2 achieved the highest observed absolute F1-score, MCC, and AUC-ROC values and greater relative stability; Dataset 1 showed intermediate performance and greater alternation among configurations; and Dataset 3 remained close to zero in MCC and close to 0.5 in AUC-ROC, so its first internal position did not indicate high absolute performance. Resampling mainly increased recall and the positive-class F1-score, without uniformly improving the other metrics. Feature selection performed on the training data did not show a consistent average improvement over the complete representation, and the effects of the reduced representations depended on the dataset and representation. Neither sex-based subgroup consistently outperformed the other. SHAP was applied to one representative run of each Scenario 1 reference configuration, rather than to an average of 30 SHAP runs. The main attributes were OverTime and JobRole in Dataset 1; satisfaction_level, number_project, and time_spend_company in Dataset 2; and Number_of_Companies_Worked, Relationship_with_Manager, and Job_Involvement in Dataset 3, with additional caution in the latter because of the model’s low discriminative ability. SHAP contributions describe predictive behavior and do not establish causality. The results do not support a universally superior model or strategy, and the observed internal stability does not replace external validation in real organizations.