Robust optimal classification trees under noisy labels

Abstract In this paper we propose a novel methodology to construct Optimal Classification Trees that takes into account noisy labels may occur in the training sample. The motivation of new is based on superaditive effect combining together margin classifiers and outlier detection techniques. Our approach rests two main elements: (1) splitting rules for classification trees are designed maximize separation between classes applying paradigm SVM; (2) some sample allowed be changed during construction tree trying detect label noise. Both features considered integrated design resulting Tree. We present Mixed Integer Non Linear Programming formulation problem, suitable solved using any available off-the-shelf solvers. model analyzed tested battery standard datasets taken from UCI Machine Learning repository, showing effectiveness our approach. computational results show most cases outperforms both accuracy AUC benchmarks provided by OCT OCT-H.