Guanqun Cao
Michigan State University
3:30 PM
154 Hurley Hall
Deep LassoNet for Functional Data Analysis
LassoNet, a neural network architecture that combines Lasso regularization withdeep learning techniques, has shown promise in feature selection for supervised learning, particularly in sparse deep neural networks for high-dimensional cross-sectional data. However, its application to functional data analysis remains relatively unexplored. In this study, we adapt the LassoNet algorithm to address the challenges specific to functional classification and functional graphical models. For functional classification, our method emphasizes the efficient selection of significant functional variables to improve both classification accuracy and model interpretability. In the context of functional graphical models, we utilize LassoNet to estimate the neighborhood of each node using an arbitrary nonparametric form, enabling the recovery of the full graph structure by combining these estimated neighborhoods. Through two real data analyses, we demonstrate LassoNet’s potential to enhance both performance and interpretability in functional data analysis.
View Poster
Originally published at acms.nd.edu.