Presentation 1-4, 4:30PM~5:00PM JST.
Speaker
Donguk Shin (Seoul National University)
Title
Foward Learning algorithm for High-dimensional Linear Structural Equation Models via l1-Regularized Regression.
Abstract
This study focuses on a forward learning approach for estimating linear structural equation models in high-dimensional settings. Although regularized methods are widely applied in backward learning algorithms, they have not yet been used in forward learning due to difficulties in controlling sparsity. Hence, this study develops a forward learning approach using L1-regularized regression and establishes its high-dimensional consistency by introducing a new concept of sparsity. The proposed algorithm is statistically consistent and computationally tractable for learning a high-dimensional sparse linear SEM, even when the maximum degree of the (moralized) graph is O(p). Various numerical experiments and real data analyses verify that the proposed algorithm is statistically consistent and computationally feasible.