An evolving feature weighting framework for radial basis function neural network models

Via Granular Computing (GrC), one can create effective computational frameworks for obtaining information from data, motivated by the human perception of combining similar objects. Combining knowledge gained via GrC with a Fuzzy inference engine (Neural-Fuzzy) enable us to develop a transparent system. While weighting variables based on their importance during the iterative data granulation process has been proposed before (W-GrC), there is no work in the literature to demonstrate effectiveness and impact on Type-2 Fuzzy Logic systems (T2-FLS). The main contribution of this paper is to extend W-GrC, for the first time, to both Type-1 and Type-2 models known as Radial Basis Function Neural Network (RBFNN) and General Type-2 Radial Basis Function Neural Network (GT2-RBFNN). The proposed framework is validated using popular datasets: Iris, Wine, Breast Cancer, Heart and Cardiotocography. Results show that with the appropriate selection of feature weight parameter, the new computational framework achieves better classification accuracy outcomes. In addition, we also introduce in this research work an investigation on the modelling structure's interpretability (via Nauck's index) where it is shown that a good balance of interpretability and accuracy can be maintained.