Causal calibration duration and online updating

Determine the minimum duration of a user-specific calibration segment recorded before deployment and characterize the effect of online updating for causal wrist-electrodermal-activity affect-recognition systems that normalize held-out users’ signals without using future samples.

Background

The paper contrasts global-train normalization, which uses statistics computed from training subjects and is compatible with inductive first-use deployment, with per-subject-z normalization, which computes statistics from the held-out user’s complete unlabeled recording. Because complete-record normalization uses samples recorded after a prediction window, it is retrospective rather than causal.

The authors note that a deployable causal system could instead use a defined segment recorded before use or update its normalization statistics online. However, the manuscript does not establish how long that initial calibration segment should be or how online updating would affect affect-recognition performance, leaving both issues unresolved for practical deployment.

References

A causal deployment would rely on training statistics or on a defined segment recorded before use; the minimum duration of such a segment and the effect of online updating remain to be studied.

— Artifact Annotations Partially Substitute for Per-User Calibration: SAFE-EDA and a Normalization-Controlled Evaluation of Wrist-EDA Affect Recognition  (2610.01692 - Chai et al., 1 Oct 2026) in Section IV-B, “Normalization Regimes and Model”