Title: An adaptive-LASSO algorithm for feature selection in functional spatiotemporal models
Authors: Paolo Maranzano - University of Milano-Bicocca & Fondazione Eni Enrico Mattei (Italy) [presenting]
Alessandro Fasso - University of Bergamo (Italy)
Philipp Otto - Leibniz University Hannover (Germany)
Abstract: A model selection algorithm based on adaptive-LASSO regularization for spatiotemporal models is discussed. In particular, we are interested in applying a penalized likelihood feature selection procedure to functional Hidden Dynamics Geostatistical Models, or f-HDGM. These models represent the phenomenon of interest using a mixed-effects structure, in which the latent component describes the spatiotemporal dynamics and the fixed-effects component models the interaction between the response variable and exogenous phenomena via linear regression. Model coefficients are shaped as continuous functions that vary across a functional domain. We focus on functional regression models based on B-spline basis functions for interpolation, where the number of free parameters is given by the order of the spline and the number of internal knots. We aim at identifying a robust procedure to select the subset of relevant spline basis functions used to model the relationships, employing a penalized likelihood algorithm and cross-validation to choose the best models. The proposed algorithm is applied to both simulated and real-world data. The empirical data concern the case study of hourly air pollutant concentrations observed during the lockdown period imposed in 2020 to address the spread of the COVID-19 virus in Northern Italy.