Time-Series Classification for Predicting Self-Reported Job Performance
Abstract
Widely available wearable devices allow for large-scale passive sensing of human behavior in real-world contexts. Previous research has established that features of heart rate are correlated with constructs related to workplace performance, such as stress and task engagement. In this study, we leverage a large-scale longitudinal study (N = 212, days = 60) to develop machine learning models to predict self-reported job performance. Using the Random Convolutional Kernel Transform (ROCKET) algorithm, which approximates convolutional layers in neural networks with low computational cost, estimates of daily heart rate, heart rate variability, physical activity, social activity, and day of the week were used to predict four job performance outcome measures. Results indicate the ROCKET algorithm had a large proportional increase in the accuracy of baseline models (48 to 121 increase), but the overall levels of accuracy for the best models remained modest (Matthews Correlation Coefficient ~ 0.10).
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 18, 2021
- Accession Number
- AD1153640
Entities
People
- Alexander F. Danvers
- Esther Sternberg
- Evan C. Carter
- Matthias R. Mehl
Organizations
- United States Army Research Laboratory