Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

WELD: A Large-Scale Longitudinal Dataset of Emotional Dynamics for Ubiquitous Affective Computing

Xiao Sun

TL;DR

WELD delivers the first large-scale, longitudinal, in-situ dataset of workplace emotions from a small team over 30.5 months, capturing 733,651 facial emotion records with seven probabilities each and augmented by 32 extended metrics, rich metadata, and pandemic-context events. By validating data quality through multiple strands (human agreement, replication of weekend/diurnal patterns, turnover prediction with AUC=$1.00$, and personality-emotion alignment), and providing strong baselines with RF, SVM, and LSTM models (emotion classification accuracy $=91.2 ext{%}$, valence $R^2=0.84$, turnover AUC=$1.00$), the work establishes a rigorous benchmark for ubiquitous affective computing in real-world settings. The dataset enables diverse research directions, including emotion recognition benchmarking, long-term emotional dynamics, contagion networks, HR analytics, and privacy-preserving, edge-enabled monitoring for small groups, all while addressing ethical and privacy considerations. This resource has practical implications for organizational wellbeing, adaptive interfaces, and proactive interventions during crises such as COVID-19, and paves the way for cross-cultural and multi-modal extensions in future work.

Abstract

Automated emotion recognition in real-world workplace settings remains a challenging problem in affective computing due to the scarcity of large-scale, longitudinal datasets collected in naturalistic environments. We present a novel dataset comprising 733,651 facial expression records from 38 employees collected over 30.5 months (November 2021 to May 2024) in an authentic office environment. Each record contains seven emotion probabilities (neutral, happy, sad, surprised, fear, disgusted, angry) derived from deep learning-based facial expression recognition, along with comprehensive metadata including job roles, employment outcomes, and personality traits. The dataset uniquely spans the COVID-19 pandemic period, capturing emotional responses to major societal events including the Shanghai lockdown and policy changes. We provide 32 extended emotional metrics computed using established affective science methods, including valence, arousal, volatility, predictability, inertia, and emotional contagion strength. Technical validation demonstrates high data quality through successful replication of known psychological patterns (weekend effect: +192% valence improvement, p < 0.001; diurnal rhythm validated) and perfect predictive validity for employee turnover (AUC=1.0). Baseline experiments using Random Forest and LSTM models achieve 91.2% accuracy for emotion classification and R2 = 0.84 for valence prediction. This is the largest and longest longitudinal workplace emotion dataset publicly available, enabling research in emotion recognition, affective dynamics modeling, emotional contagion, turnover prediction, and emotion-aware system design.

WELD: A Large-Scale Longitudinal Dataset of Emotional Dynamics for Ubiquitous Affective Computing

TL;DR

WELD delivers the first large-scale, longitudinal, in-situ dataset of workplace emotions from a small team over 30.5 months, capturing 733,651 facial emotion records with seven probabilities each and augmented by 32 extended metrics, rich metadata, and pandemic-context events. By validating data quality through multiple strands (human agreement, replication of weekend/diurnal patterns, turnover prediction with AUC=, and personality-emotion alignment), and providing strong baselines with RF, SVM, and LSTM models (emotion classification accuracy , valence , turnover AUC=), the work establishes a rigorous benchmark for ubiquitous affective computing in real-world settings. The dataset enables diverse research directions, including emotion recognition benchmarking, long-term emotional dynamics, contagion networks, HR analytics, and privacy-preserving, edge-enabled monitoring for small groups, all while addressing ethical and privacy considerations. This resource has practical implications for organizational wellbeing, adaptive interfaces, and proactive interventions during crises such as COVID-19, and paves the way for cross-cultural and multi-modal extensions in future work.

Abstract

Automated emotion recognition in real-world workplace settings remains a challenging problem in affective computing due to the scarcity of large-scale, longitudinal datasets collected in naturalistic environments. We present a novel dataset comprising 733,651 facial expression records from 38 employees collected over 30.5 months (November 2021 to May 2024) in an authentic office environment. Each record contains seven emotion probabilities (neutral, happy, sad, surprised, fear, disgusted, angry) derived from deep learning-based facial expression recognition, along with comprehensive metadata including job roles, employment outcomes, and personality traits. The dataset uniquely spans the COVID-19 pandemic period, capturing emotional responses to major societal events including the Shanghai lockdown and policy changes. We provide 32 extended emotional metrics computed using established affective science methods, including valence, arousal, volatility, predictability, inertia, and emotional contagion strength. Technical validation demonstrates high data quality through successful replication of known psychological patterns (weekend effect: +192% valence improvement, p < 0.001; diurnal rhythm validated) and perfect predictive validity for employee turnover (AUC=1.0). Baseline experiments using Random Forest and LSTM models achieve 91.2% accuracy for emotion classification and R2 = 0.84 for valence prediction. This is the largest and longest longitudinal workplace emotion dataset publicly available, enabling research in emotion recognition, affective dynamics modeling, emotional contagion, turnover prediction, and emotion-aware system design.
Paper Structure (57 sections, 10 equations, 4 figures, 6 tables)

This paper contains 57 sections, 10 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Motivation
  3. Existing Datasets and Limitations
  4. Contributions
  5. Ethical Considerations
  6. Related Work
  7. Facial Expression Datasets
  8. Workplace Emotion Measurement
  9. Longitudinal Emotion Studies
  10. COVID-19 and Workplace Emotions
  11. Dataset Description
  12. Participants and Setting
  13. Temporal Coverage and Sampling
  14. Facial Expression Recognition System
  15. Facial Expression Recognition System
  16. ...and 42 more sections

Figures (4)

  • Figure 1: Dataset Overview. (A) Temporal coverage illustrating data collection density over 30.5 months, with key COVID-19 events highlighted. (B) Distribution of record counts per participant, differentiated by employment status (blue: active, red: departed). (C) Emotion distribution across seven fundamental emotions. (D) Role distribution spanning six job categories.
  • Figure 2: Extended Emotional Metrics. (A) Scatter plot of volatility versus predictability showing a negative correlation ($r=-0.58$). (B) Distribution of emotional inertia with a median of 0.42. (C) Relationship between baseline valence and volatility, indicating higher volatility among individuals with negative baseline valence. (D) Comparison of emotion ratios between active and departed employees, revealing elevated fear and anger in departed individuals.
  • Figure 3: Emotion Patterns and Technical Validation. (A) Valence distribution exhibiting the negative skew typical of workplace stress. (B) Diurnal rhythm with valence lowest at work start (9--10 AM) and peaking at lunch (12--1 PM). (C) Weekend effect showing a 192% valence increase on weekends ($p<0.001$). (D) Emotion correlation matrix revealing expected relationships (happy--sad anticorrelation: $r=-0.62$).
  • Figure 4: Individual emotional trajectories for 12 representative participants over 30.5 months. Each line shows the daily average valence score (smoothed with 7-day moving average) for one participant. The vertical shaded region indicates the Shanghai lockdown period (April-June 2022). Participants were selected to illustrate the diversity of emotional patterns in the dataset, including differences in baseline valence, temporal stability, and responsiveness to external events.