Mejora de la estimación del esfuerzo en proyectos de software mediante métodos de sobremuestreo y aprendizaje computacional

Effort estimation prediction determines the time it will take to develop a software program or the resources required to complete it within the established timeframe. A current alternative for predicting estimates is to use machine learning methods. However, publicly available data sets generally contain few samples, so such methods cannot improve their effectiveness. Thus, it is necessary to increase the number of samples using oversampling methods. Therefore, this paper presents the use of ensemble methods with combinations of oversampling and undersampling to analyze the performance impact of the regressors used on small and medium-sized data sets. Moreover, their effectiveness in improving effort estimation in software projects using measures such as MMRE, MAE, RMSE, and Pred is also presented. The results obtained from MMRE and Pred, mainly show that the application of these strategies reduces prediction errors. Consequently, the use of an appropriate ensemble model, together with oversampling and undersampling strategies, improves effort prediction, especially on small data sets such as COCOMO, Maxwell, and Desharnais with highly unbalanced sample distributions.