Progressive granular neural networks with class based granulation

Abstract

Granular neural networks (GNNs) process granulated data with neural networks. Class based (CB) granulation of input data considers the belongingness of each feature to the classes present in the data. Advancement in the sensor technology has produced large amount of data which is also called as stream of data. To process the stream of CB granulated data, the article proposes an adaptive neural network model which is named as class based progressive granular neural network (CBPGNN). Learning or updating weights in CBPGNN is based on back propagating the output error. This overcomes the disadvantage of evolving granular neural networks (EGNNs) in which updation of weights is done linearly. CBPGNN posses the advantages of both class based granulation and progressive granular neural network (PGNN). The performance of CBPGNN is tested on remote sensing and waveform datasets. The superiority of proposed model has been justified by comparing its performance with similar types of models. The performance of proposed model is measured using the indices like average accuracy and kappa coefficient.