On the fusion of neural network models in the case of heteroscedasticity

Abstract

In this study, a neural network based clustering algorithm by robust measure was developed with the intention of filtering outliers from a given dataset. Outliers in data set usually make a model to deviate from the assumption of homoscedastic relationship which in turn leads to a heteroscedastic relationship. The developed clustering based algorithm uses two types of neural network. i.e. Cascade forward backpropagation neural network and feed forward neural network. The clustering algorithm was based on the robust estimates of location and dispersion matrix. Five (5) independent data sets obtained from the UCI machine learning repository data link were used for this study. Two techniques i.e. CFBNFDCARM and FFNNFDCARM were employed with the aim of comparing the performance of each of the technique. The evaluating metrics of mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE) and the mean absolute percentage error (MAPE) were used as the measure of the performance function in this study. From the obtained results, CFBNFDCARM technique came out to be better than the FFNNFDCARM technique. The analyses of this study were simulated using both MATLAB R2014 software and R i386 version 3.3.0 software. The performance of each technique were further plotted in order to illustrate the findings of this study.