Combining neural-based regression predictors using an unbiased and normalized linear ensemble model

In this paper, we combined a group of local regression predictors using a novel unbiased and normalized linear ensemble model (UNLEM) for the design of multiple predictor systems. In the UNLEM, the optimization of the ensemble weights is formulated equivalents to a constrained quadratic programming problem, which can be solved with the Lagrange multiplier. In our simulation experiments of data regression, the proposed multiple predictor system is composed of three different types of local regression predictors, and the effectiveness evaluation of the UNLEM was carried out on eight synthetic and four benchmark data sets. Results of the UNLEM's performance in terms of mean-squared error are significantly lower, in comparison with the popular simple average ensemble method. Moreover, the UNLEM is able to provide the regression predictions with a relatively higher normalized correlation coefficient than the results obtained with the simple average approach.