1284-GRM HAR-IQ: Training-Data Integrity Assessment and Selective Recovery for Robust Wearable Human Activity Recognition

Presenter Information

Location

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Document Type

Event

Start Date

20-11-2026 2:00 PM

Description

Wearable human activity recognition (HAR) uses sensor measurements to identify human activities, but incorrect training labels can affect recognition performance. This ongoing research investigates training-data integrity and selective recovery to improve the reliability of wearable HAR. Preliminary experiments use the KU-HAR dataset with random and structured training-label corruption at rates of 10%, 20%, and 30%. Across 18 experimental runs, suspicious labels are identified using out-of-fold predictions, and different correction strategies are evaluated. At 20% random corruption, detection achieved a mean average precision of 0.944. However, selective correction did not improve mean recognition performance across the evaluated settings. These findings highlight the need for safer recovery strategies and further evaluation.

This document is currently not available here.

Share

COinS
 
Nov 20th, 2:00 PM

1284-GRM HAR-IQ: Training-Data Integrity Assessment and Selective Recovery for Robust Wearable Human Activity Recognition

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Wearable human activity recognition (HAR) uses sensor measurements to identify human activities, but incorrect training labels can affect recognition performance. This ongoing research investigates training-data integrity and selective recovery to improve the reliability of wearable HAR. Preliminary experiments use the KU-HAR dataset with random and structured training-label corruption at rates of 10%, 20%, and 30%. Across 18 experimental runs, suspicious labels are identified using out-of-fold predictions, and different correction strategies are evaluated. At 20% random corruption, detection achieved a mean average precision of 0.944. However, selective correction did not improve mean recognition performance across the evaluated settings. These findings highlight the need for safer recovery strategies and further evaluation.