# Enabling Older Adults to Provide High-quality Activity Labels: Unpacking Accuracy, Precision, and Granularity in Activity Labeling

## TL;DR

This paper explores how older adults can label their activities effectively to train personalized trackers, balancing accuracy and user burden.

## Contribution

The study introduces user-centered strategies for activity labeling that reduce intrusiveness while maintaining label quality.

## Key findings

- Participants preferred user-initiated labeling to maintain control and reduce interruptions.
- Machine-initiated prompting was seen as helpful but needed to be carefully timed to avoid disruption.
- Thematic analysis revealed discrepancies between user perceptions and technical standards in labeling.

## Abstract

High-quality labels of activity data with broad representations and real-world variability are key to developing activity recognition models tailored to the needs and characteristics of older adults. However, labeling real-world data presents significant challenges, placing a heavy burden on users to provide high-quality labels while staying engaged in their activities. This paper investigates older adults’ perceptions of providing high-quality labels in the context of training their personalized activity trackers. We conducted a co-design study with 12 older adults to envision the labeling process—describing activity names and time spans—using the teachable machines paradigm as a scaffold. We unpack the contextualized definitions of accuracy, precision, and granularity through a thematic analysis of older adults’ perspectives on activity labeling. Our findings present participants’ preferred strategies for obtaining high-quality activity labels with less burden and intrusiveness, including user-initiated labeling and machine-initiated prompting. We discuss design considerations for future data labeling tools that address discrepancies between user perceptions and technical standards in training personalized activity trackers.