Resumen de Unified Performance Measure for Binary Classification Problems

Ana R. Redondo, Jorge Navarro, Rubén R. Fernández, Isaac Martín de Diego , Javier Martínez Moguerza , Juan José Fernández Muñoz

• Different performance measures are used to inspect, compare and evaluate the behaviour of classifiers in Machine Learning (ML). ML researchers employ one or several of these performance measures in their classification studies to report their success. However, no widespread consensus has been reached on a unified chosen measure. In this work, we introduce a reliable and informative measure, the Unified Performance Measure (UPM), by modifying the F1-score to avoid its undesired behaviour in imbalanced classification problems. UPM is compared with alternative performance measures, like the F1-score or Accuracy, in both simulated confusion matrices and real datasets. The proposed measure outperforms the alternatives, providing a promising new research line.