Finding Originally Mislabels with MD-ELM
Anton Akusok, David Veganzones, Yoan Miché, Éric Séverin, Amaury Lendasse · 2014
Abstract. This paper presents a methodology which aims at detecting mislabeled samples, with a practical example in the field of bankruptcy prediction. Mislabeled samples are found in many classification problems and can bias the training of the desired classifier. This paper proposes a new method based on Extreme Learning Machine (ELM) which allows for identification of the most probable mislabeled samples. Two datasets are used in order to validate and test the proposed methodology: a toy example (XOR problem) and a real dataset from corporate finance (bankruptcy prediction). 1