A Randomized Ensemble Based on Distinguishing Output Labellings]{A Randomized Ensemble Based on Distinguishing Output Labellings

This article proposes and analyzes a novel classification model trained to distinguish correctly from incorrectly labeled instances. The proposed method transforms any classification problem from a task of producing the correct label to a task of deciding if a given label is correct. To do so, a new dataset is created in which an attribute is added with a randomized label suggestion for each instance. The model then learns if the suggested labels are correct or not.The method builds an ensemble of decision trees trained on different randomized versions of this classification task.The proposed ensemble is validated with several experiments with the aim of giving some insights on the algorithm performance, especially its hyper-parameters, and to compare its performance with benchmark models. The experiments, carried on a variety of datasets, real-world and synthetic,shows that the proposed model achieves competitive results with respect to state-of-the-art learning models of the literature such as Random Forest.