Nonlinear Support Vector Regression Methodology for Modelling and Prediction: An Application

Abstract

The main limitation of Multiple linear regression analysis for estimating cause-effect relationship is highlighted. Artificial neural network (ANN) methodology that does not require specification of exact nonlinear functional relationship between a response and a set of predictor variables is briefly discussed. Some advantages and disadvantages of this technique are pointed out. The recently developed Nonlinear support vector regression (NLSVR) methodology, which is very promising and versatile, is described. As an illustration, Maize crop yield data as response variable and Total human labour, Farm power, Fertiliser consumption and Pesticide consumption as predictor variables are considered. Both ANN and NLSVR techniques for modelling and prediction purposes are employed. Performance of a fitted model is assessed in terms of Root mean square error (RMSE), Mean absolute error (MAE) and Mean absolute prediction error (MAPE). STATISTICA software package is used for carrying out data analysis. Superiority of NLSVR technique over ANN technique is showed for the data under consideration. It is concluded that NLSVR methodology is quite successful for modelling as well as prediction purposes