Multiclass Classification using Neural Networks and Interval Neutrosophic Sets

This paper presents a new approach to the problem of multiclass classification. The proposed approach has the capability to provide an assessment of the uncertainty value associated with the results of the prediction. Two feed-forward backpropagation neural networks, each with multiple outputs, are used. One network is used to predict degrees of truth membership and another network is used to predict degrees of false membership. Indeterminacy membership or uncertainty in the prediction of these two memberships is also estimated. Together these three membership values form an interval neutrosophic set. Hence, a pair of single multiclass neural networks with multiple outputs produces multiple interval neutrosophic sets. We experiment our technique to the classical benchmark problems including balance, ecoli, glass, lenses, wine, yeast, and zoo from the UCI machine learning repository. Our approach improves classification performance compared to an existing technique which applied only to the truth membership created from a single neural network with multiple outputs. Key–Words:multiclass classification, uncertainty, interval neutrosophic sets, multiclass neural network, feedforward backpropagation neural network