Longer Records, Broader Invariance: The Hidden Scaling Problem in Longitudinal Contrastive Learning

2026-09-28

AI Summary

This study addresses a critical limitation in longitudinal contrastive learning, where expanding the recording span leads to an excessively broad supervision scope. The resulting over-abundance of positive pairs induces invariance bias, erasing key information about individual temporal dynamics. To overcome this, we formally disentangle the concepts of recording span and supervision span for the first time, proposing a person-level contrastive learning framework that explicitly contrasts other observations from the same individual to reverse information loss. Prospective validation using home sensing data and the GLOBEM cohort demonstrates that our approach effectively mitigates the suppression of change-related signals caused by broad supervision. By significantly recovering individual dynamic features while preserving the advantages of extended recordings, this work establishes a new paradigm for longitudinal representation learning.

Abstract

Longitudinal data are valuable because people change. Yet the objectives used to learn from these data can inadvertently erase that change. In person-level contrastive learning, observations from the same person are treated as positives; as records grow, those positives can span increasingly distant---and increasingly different---behavioral states. More history can therefore produce not only more data, but broader invariance. We show that this distinction is fundamental. We separate \emph{record span}, how much history the learner sees, from \emph{supervision span}, how far across that history positive-pair supervision reaches. Across in-home sensing records spanning up to 2.7 years, broader supervision systematically suppresses recoverable changing-state information, even when the available history is held fixed. At the broadest span, less than 10\% of the information recoverable from an untrained encoder remains. Yet keeping positives local is not sufficient: as records grow, even distant states that are never paired become increasingly similar. Explicitly contrasting other observations from the same person reverses this loss without shortening the record, revealing a second route by which longitudinal scale can broaden invariance. Finally, we prospectively reproduce the supervision-span effect in 199 GLOBEM participants. Longitudinal scale therefore presents a choice: more history need not mean more invariance. By controlling what is held invariant as records grow, we can preserve the change that made the longitudinal data valuable in the first place.

Problem

• longitudinal contrastive learning
• invariance
• record span
• supervision span
• representation learning

Innovation

• Longitudinal Contrastive Learning
• Supervision Span
• Record Span
• Invariance
• Representation Learning