3WDNFS – Three-way decision neuro-fuzzy system for classification

Abstract

The three-way decision approach is an emerging paradigm in the design of tools for data mining and machine learning. It switches from a two-way classification (“negative” and “positive” class) to three decisions: “negative”, “positive”, and non-commitment class. It means that when for some data it is not possible to elaborate a reliable answer they are assigned with a non-commitment class. In the paper we apply this paradigm for a cascade of neuro-fuzzy classifiers. If the first neuro-fuzzy system assigns a data item with a non-commitment class, the next neuro-fuzzy system is run for this data item. For easy items the first system is enough, but for harder ones two or more systems have to be run. Neuro-fuzzy systems elaborate interpretable fuzzy models. The models are composed of fuzzy rules that can be interpreted linguistically by humans. Application of neuro-fuzzy systems results in a cascade of interpretable models. The paper describes algorithms for training a cascade of neuro-fuzzy classifiers and for elaboration of answers. The paper presents results of numerical experiments that show that this technique can elaborate results with lower generalisation error than two-way classifiers. The implementation of the proposed system is available from a github repository.