Stepwise Linear Regression for Dimensionality Reduction in Neural Network Modelling

This work considers the applicability of applying the derivatives of stepwise linear regression modelling (specifically the p-values which indicate the importance of a variable to the modelling process) as a feature extraction technique. We utilise it in conjunction with several data sets of varying levels of complexity, and compare our results to other dimensionality reduction techniques such as genetic algorithms, sensitivity analysis and linear principal components analysis prior to data modelling using several different neural network models. Our results indicate that stepwise linear regression is highly effective in this role with results comparable to and sometimes superior then more established techniques