AI
Researchers Release First Long-Term Workplace Emotion Dataset Spanning 30 Months
A new Chinese workplace study captures nearly 750,000 facial expressions over two and a half years, revealing patterns in employee emotions and exposing AI bias in facial recognition systems.
Researchers have created WELD, a novel emotion-recognition dataset that tracks facial expressions across an entire workplace over an extended period — a breakthrough that fills a significant gap in affective computing research.
What Makes WELD Unique#
The dataset combines four features rarely found together in emotion research: it spans 30.1 months (November 2021 through May 2024), was collected in a real workplace setting rather than a controlled lab, tracks a stable group of 49 employees at a Chinese software company, and uses entirely passive sensing with full institutional approval. The dataset contains 733,780 seven-category facial-expression measurements, making it the longest naturalistic emotion corpus available and the only multi-year resource that enables both individual-level longitudinal tracking and team-based relational analysis.
To protect privacy, WELD is released through a tiered access model with only aggregated probability data publicly available.
Key Findings on Workplace Emotions#
The researchers validated their corpus by confirming three previously documented patterns: a 43.1% boost in positive emotion on weekends, a trough in emotional positivity around 1 p.m. each day, and measurable emotional impacts from Shanghai's 2022 pandemic lockdown (effect size d=-0.40).
Beyond validation, four novel discoveries emerged from the data:
- Emotional variance patterns: Individual differences account for 19.3% of day-to-day mood swings, while seasonal monthly patterns explain 29.8% — establishing a quantitative upper bound for how well future predictive models might perform
- Emotional regimes: Hidden Markov analysis uncovered six distinct emotional states, with negative emotions persisting far longer (16-18 days) than positive ones (3 days)
- Turnover prediction limits: A leave-one-person-out model achieved 0.79 AUC in predicting employee departures, yet a Cox concordance index revealed only 0.52 performance — highlighting how standard metrics can overstate predictive power when survival analysis isn't considered
- Facial recognition bias: Commercial facial recognition systems substantially over-identify anger on neutral Asian faces (0.194 vs. ~0.05 based on Western training data), making WELD valuable for auditing fairness in emotion AI
