Avaliação comparativa de técnicas de aprendizado de máquina na previsão de Turnover, com uso de inteligência artificial explicável.

Abstract

This article carried out a comparative analysis of ten Machine Learning algorithms for predicting turnover scenarios in organizations. Two databases from the Kaggle website were used: the first for an American company, whose name was not disclosed, with 9,540 records and the other for Portobello Tech's employee turnover prevention work with 11,995 records, both with ten attributes each. The metrics cataloged were Precision, F1-Score, Recall and Accuracy. Given the results of the algorithms, Random Forest stood out, with an accuracy of 0.84 for the first base and 0.95 for the second in relation to the test sets of the bases used. After this stage, the SHAP methodology was used to identify the attributes with the greatest impact on turnover intention. The average number of hours worked per month, employee satisfaction and time with the company were the most relevant attributes in both bases for the decision to resign.