Skip to content

CV criterions and tuning #20

Description

@erikcs

Some "heuristic" covariate imbalance related CV criterions are

  • balance loss (Tan, 2020)
  • average norm of gradient of balance loss = norm of covariate imbalance (Zhao, 2019)
  • mean covariate imbalance, SMD, (Wang & Zubizarreta, 2020; Chattopadhyay, Hase, & Zubizarreta, 2020)

AFAIK these CV criterions are not explicitly theoretically grounded, are some open questions in the literature.

This paper discusses alternative ways of selecting the level of regularization.

(Ensure: do not divide by zero as all these metrics are evaluated out-of-fold)

  • For augmentation/debiasing end-use cases, there's the Riesz loss.

Metadata

Metadata

Assignees

No one assigned

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions