Aki Vehtari, PhD Associate ProfessorDepartment of Computer SciencesAalto University
Variable selection is often not needed, and a big model with all potentially
relevant variables can be used. Such a big model can avoid overfitting if
appropriate priors are used. Variable selection can be useful to reduce future
measurement costs or improve explainability. If we have built a good big model
with all variables, that can be used as a reference model, and we can condition
the variable selection decision task on that reference model. Reference model
approach greatly improves the stability of variable selection. I present the
basic ideas and some recent advances in the reference model variable selection
approach.