Information-theoretic approach to variable selection in predictive models applied to paper machine data

This paper presents an information-theoretic approach to variable selection for prediction of laboratory measurements of paper quality. Along with a well-known Principal Component Analysis we considered techniques for variable selection based on the classical Shannon Mutual Information and a novel Maximal Information Coefficient. A multilayer perceptron neural model was used to predict quality measurements and compare feature selection techniques. The suggested approach was tested on real industrial data obtained form a pilot paper machine. The presented results show that information-theoretic techniques perform better compared to Principal Component Analysis, providing higher accuracy results.