Feature Selection and Multivariate Regression for Paper Surface Roughness Prediction

Abstract

Paper surface roughness is an important quality parameter because it affects coating quality, printability, and the final appearance of paper. Accurate prediction of roughness remains challenging due to the large number of correlated variables monitored throughout the papermaking process. This study compares three feature selection methods for predicting average paper roughness from industrial data: Pearson correlation analysis, Principal Component Analysis (PCA), and a hybrid approach combining the Least Absolute Shrinkage and Selection Operator (Lasso) with Forward Selection. The dataset consisted of 7,145 hourly observations and 198 process variables collected from an industrial paper production process. After data preprocessing and min-max normalization, the selected variables were used to develop Multiple Linear Regression (MLR) models. The Pearson-based model achieved an R² of 0.696, whereas the PCA model reached an R² of 0.728 using seven principal components that explained 92.37% of the data variance. The best performance was obtained with the Lasso + Forward Selection approach, which achieved an R² of 0.789 using only 20 original process variables. Although PCA reduced the dimensionality of the dataset, the hybrid approach produced a more accurate model while preserving variables with direct physical meaning. These results indicate that combining Lasso with Forward Selection is an effective strategy for predicting paper roughness and can support process monitoring and quality improvement in industrial papermaking.