TLDR: The WELD dataset is the largest and longest-running publicly available longitudinal dataset of workplace emotions, comprising over 733,000 facial expression records from 38 employees over 30.5 months. It captures seven emotion probabilities, rich metadata, and 32 advanced emotional metrics, including responses to the COVID-19 pandemic. The dataset validates known psychological phenomena like weekend mood boosts and diurnal rhythms, flawlessly predicts employee turnover, and aligns with personality traits. It serves as a crucial resource for advancing ubiquitous affective computing, emotion recognition, and organizational behavior research.
Understanding human emotions in real-world settings has always been a significant challenge for affective computing, the field dedicated to building systems that can recognize, interpret, and simulate human emotions. Most existing datasets are either collected in controlled lab environments with posed expressions or are too short to capture long-term emotional patterns. To bridge this critical gap, researchers have introduced WELD: a groundbreaking, large-scale longitudinal dataset designed to capture the intricate emotional dynamics within a genuine workplace.
The WELD dataset is an unprecedented resource, featuring 733,651 facial expression records collected from a small team of 38 employees over an extensive period of 30.5 months, from November 2021 to May 2024. This makes it the largest and longest-running publicly accessible longitudinal dataset of workplace emotions. Each record in WELD includes probabilities for seven fundamental emotions: neutral, happy, sad, surprised, fear, disgusted, and angry. Beyond these basic emotions, the dataset is enriched with comprehensive metadata such as job roles, employment outcomes (including a 40.7% turnover rate), and personality traits (Big Five).
One of WELD’s most compelling features is its temporal coverage, spanning the challenging period of the COVID-19 pandemic. This allows researchers to analyze emotional shifts linked to major societal events, such as the Shanghai lockdown (April–June 2022) and evolving policy changes. To provide an even deeper understanding of emotional states, the dataset further includes 32 advanced emotional metrics derived from established affective science frameworks. These metrics encompass concepts like valence (pleasantness), arousal (activation), volatility (emotional instability), predictability, inertia (emotional persistence), and emotional contagion strength, offering a nuanced view of individual emotional experiences.
The technical validation of WELD confirms its robustness and reliability. The dataset successfully replicates well-known psychological phenomena, such as a significant 192% valence boost on weekends (indicating improved mood away from work) and consistent diurnal rhythms in emotional states. It also objectively quantifies the emotional impact of the COVID-19 pandemic, showing a valence drop during lockdown and a surge upon reopening. Remarkably, WELD demonstrates flawless predictive power for employee turnover, achieving an AUC of 1.0, suggesting that emotional patterns can strongly indicate future employment decisions. Furthermore, the dataset shows a strong alignment between long-term emotional patterns and independently rated personality traits.
Baseline models, including Random Forest and Long Short-Term Memory (LSTM) networks, have been applied to the dataset, achieving impressive results. The LSTM model, in particular, reached 91.2% accuracy in emotion classification and an R-squared value of 0.84 for valence prediction, highlighting the value of temporal modeling in understanding emotions. These results set a strong foundation for future research in emotion recognition and affective computing.
WELD opens up numerous new avenues for research. It serves as a benchmark for emotion recognition in naturalistic, ubiquitous environments, enabling studies on long-term emotional tracking, the dynamics of emotional contagion within small groups, and the forecasting of employee turnover. It also supports the design of emotion-aware systems and ubiquitous psychological services that can unobtrusively monitor and support employee well-being. The dataset is freely available under a CC BY 4.0 license, encouraging broad scientific collaboration.
While the dataset offers profound insights, the researchers have also meticulously addressed ethical considerations. All participants provided informed consent, and stringent privacy protection measures were implemented. These include on-site processing of video data (only aggregated emotion probabilities are stored, no raw video), pseudonymization of participant IDs, data minimization (no direct identifiers or demographics), and strict ethical use restrictions to prevent misuse for surveillance or discriminatory purposes. This commitment to privacy and ethical use is paramount for advancing ubiquitous affective computing responsibly.
Also Read:
- Advancing Parkinson’s Assessment with Relaxed Multimodal AI
- Unlocking Trust: The Power of Interactive AI Explanations
This comprehensive dataset is poised to accelerate advancements in understanding workplace emotions and developing more sophisticated, context-aware affective computing technologies. For more detailed information, you can refer to the full research paper: WELD: A Large-Scale Longitudinal Dataset of Workplace Emotional Dynamics for Ubiquitous Affective Computing.
