Fused LassoNet: Sequential feature selection for spectral data with neural networks

Abstract

Feature selection for high-dimensional spectral data is critical to improve the accuracy and interpretability of chemometric models. Various methods for feature selection have been introduced in chemometrics; however, achieving explainable sequential feature selection while conducting nonlinear classification simultaneously remains challenging. To address the challenge, this study proposes a fused least absolute shrinkage and selection operator network (LassoNet) that integrates the regularization principles of both the LassoNet and fused Lasso within the framework of a neural network. Further, the fused Lasso method facilitates continuous feature selection by considering the sequence between features, whereas LassoNet method enables nonlinear modeling using neural networks. We solve the fused LassoNet problem with proximal gradient descent, and the optimality of the proximal operator is mathematically proved. This study analyzes the performances of Lasso, fused Lasso, LassoNet, and fused LassoNet in classifying two groups using nine spectral datasets. The fused LassoNet demonstrates superior performance in terms of classification accuracy and sequential feature selection. These results demonstrate the proposed method enhances the predictive accuracy and interpretability of chemometric models using spectral data.