Quantification of Vagueness in Multiclass Classification Based on Multiple Binary Neural Networks

This paper presents an innovative approach to solve the problem of multiclass classification. One-against-one neural networks are applied to interval neutrosophic sets (INS). INS associates a set of truth, false and indeterminacy membership values with an output. Multiple pairs of the truth binary neural network and the false binary neural network are trained to predict multiple pairs of the truth and false membership values. The difference between each pair of truth and false membership values is considered as vagueness in the classification and formed as the indeterminacy membership value. The three memberships obtained from each pair of networks constitute an interval neutrosophic set. Multiple interval neutrosophic sets are then created and used to support decision making in multiclass classification. We have applied our technique to three classical benchmark problems including balance, wine, and yeast from the UCI machine learning repository. Our approach has improved classification performance compared to an existing one-against-one technique which applies only to the truth membership values.