Three–Way Classification: Ambiguity and Abstention in Machine Learning

Ambiguity, that is the lack of information to produce a specific classification, is an important issue in decision–making and supervised classification. In case of ambiguity, human–decision makers can resort to abstaining from making precise classifications (especially when error-related costs are high), but this behaviour has been scarcely addressed, and applied, in machine learning algorithms. This contribution grounds on previous works in the areas of three–way decisions, cautious classification and orthopairs, and proposes a set of techniques we developed to address this form of ambiguity, by providing both a general–purpose technique to create three–way algorithms from probabilistic ones, and also more specific techniques which could be applied to popular machine learning frameworks. We also evaluate the proposed idea, by performing a set of experiments where we compare classical classification algorithms with the corresponding three–way generalizations, in order to study the trade–off between classification accuracy and abstention: the results are promising.