摘要：Many classification systems rely on clustering techniquesin which a collection of training examples is provided as aninput, and a number of clusters c1, . . . cmmodelling someconcept C results as an output, such that every cluster ciislabelled as positive or negative. Given a new, unlabelledinstance enew, the above classification is used to determineto which particular cluster cithis new instance belongs. Insuch a setting clusters can overlap, and a new unlabelledinstance can be assigned to more than one cluster with con-.icting labels. In the literature, such a case is usually solvednon-deterministically by making a random choice. This pa-per presents a novel, hybrid approach to solve this situationby combining a neural network for classification along witha defeasible argumentation framework which models pref-erence criteria for performing clustering
