Dealing with contaminated datasets: An approach to classifier training
Władysław Homenda, Agnieszka Jastrzębska, Mariusz Rybnik · AIP conference proceedings · 2016
The paper presents a novel approach to classification reinforced with rejection mechanism. The method is based on a two-tier set of classifiers. First layer classifies elements, second layer separates native elements from foreign ones in each distinguished class. The key novelty presented here is rejection mechanism training scheme according to the philosophy “one-against-all-other-classes”. Proposed method was tested in an empirical study of handwritten digits recognition.