A hybridized neural network models for data with outliers

Abstract

In this research, a hybridized neural network technique was developed with the aim of investigating and comparing its performance with the performance of the standalone neural network in the case of deviation from the assumption of homoscedastic relationship in dataset. Modeling using real data as asserted by [14] is presumed to contain between 1% to 10% contaminations. Presence of outliers in a dataset may contradict the assumption of normality or even both normality and homoscedasticity [9]. Two neural networks i.e Cascade forward backpropagation neural network (CFBNN) and Feed forward backpropagation neural network (FFBNN) were considered in this research. The functional potentiality of neural network has led to a verse studies comparing its performance in the predictive capability [13]. A clustering algorithm based on robust measure were introduced to each of this neural network to form a hybridized neural networks known as the cascade forward backpropagation neural network over a filtered data by clustering algorithm based on robust measure acronym as (CFBNFDCARM) and Feed forward backpropagation neural network over a filtered data by clustering algorithm based on robust measure acronym as (FFBNFDCARM). The proposed hybridized techniques were employed on six (6) different dataset obtained from data repository dataset (UCI). The clustering technique tends to filter out the outliers from each of the obtained dataset. The filtered dataset were then introduced to the two neural networks in order to determine their performances. The results obtained from the proposed hybridized neural network techniques were compared with the results obtained from the standalone neural network techniques. The comparison indicates that, the emerging performance results from the proposed hybridized techniques generally on the average outperformed the performance results from the standalone techniques in terms of the evaluating metrics of the mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) as well as the mean absolute percentage error (MAPE).